NBER WORKING PAPER SERIES

CULTURAL PROXIMITY AND LOAN OUTCOMES

Raymond Fisman Daniel Paravisini Vikrant Vig

Working Paper 18096 http://www.nber.org/papers/w18096

NATIONAL BUREAU OF ECONOMIC RESEARCH 1050 Massachusetts Avenue Cambridge, MA 02138 May 2012

Thanks to Andres Lieberman and Sravya Mamadiana for outstanding research assistance. We would like to thank Abhijit Banerjee, Oriana Bandiera, Shawn Cole, Esther Duflo, Michael Kremer, Raghu Rajan, Antoinette Schoar, Paola Sapienza, Luigi Zingales, and the participants at the CFS-EIEF Conference on Household Finance, Columbia University, Boston University, European Bank for Reconstruction and Development, London Business School, London School of Economics - Development, MIT-Sloan, the NBER conferences in corporate finance and household finance, Stockholm School of Economics, and Vienna University of Economics and Business for valuable feedback. The views expressed herein are those of the authors and do not necessarily reflect the views of the National Bureau of Economic Research.

NBER working papers are circulated for discussion and comment purposes. They have not been peer- reviewed or been subject to the review by the NBER Board of Directors that accompanies official NBER publications.

© 2012 by Raymond Fisman, Daniel Paravisini, and Vikrant Vig. All rights reserved. Short sections of text, not to exceed two paragraphs, may be quoted without explicit permission provided that full credit, including © notice, is given to the source. Cultural Proximity and Loan Outcomes Raymond Fisman, Daniel Paravisini, and Vikrant Vig NBER Working Paper No. 18096 May 2012 JEL No. D82,G21,J15

ABSTRACT

We present evidence that shared codes, religious beliefs, ethnicity - cultural proximity - between lenders and borrowers improves the efficiency of credit allocation. We identify in-group preferential treatment using dyadic data on the religion and caste of bank officers and borrowers from a bank in India, and a rotation policy that induces exogenous matching between officers and borrowers. Cultural proximity increases lending on both intensive and extensive margins and improves repayment performance, even after the in-group officer is replaced by an out-group one. Further, cultural proximity increases loan dispersion and reduces loan to collateral ratios. Our results imply that cultural proximity mitigates informational problems that adversely affect lending, which in turn relaxes financial constraints and improves access to finance.

Raymond Fisman Vikrant Vig School of Business London Business School Columbia University Regent's Park 622 Uris Hall London NW1 4S, UK 3022 Broadway [email protected] New York, NY 10027 and NBER [email protected]

Daniel Paravisini Columbia University Graduate School of Business 3022 Broadway, Uris Hall 416 New York, NY 10027 and NBER [email protected] 1 Introduction

Shared codes, language, religion, or culture between potential parties of a transaction can affect the likelihood that the transaction takes place, and also its outcome. Commonali- ties in religion and in ethnic origin, for example, are positively associated with trade flows between countries (Guiso, Sapienza and Zingales, 2009). The effect on efficiency of such cultural proximity between transacting parties is ambiguous, however. On the one hand, if members of a group tend to do business with one another for preference-based reasons, this may lead to discrimination or favoritism and result in the misallocation of resources. Alternatively, if cultural proximity reduces the cost of communication or contract enforce- ment, in-group transactions may be more efficient. Given these opposing views, the effect of cultural proximity on loan contracting remains an empirical question, something that we explore in this paper. There are a number of challenges in empirically identifying the extent of preferential in-group treatment, and distinguishing among the various explanations underlying such behavior. First, it requires information on the group membership of both transacting parties. Most studies have relied exclusively on the religion or race of only one side of the market, and have thus been best set up to detect discrimination against minorities rather than dyadic preferences for one’s own type. This confounds any beneficial effect of in-group interactions with statistical or animus-based discrimination, especially when the in-group advantages are more prevalent within relatively small minority groups. Second, even when dyadic data are available, matching between parties is driven by the transac- tions’ expected profitability, which is not observed by the econometrician. Unobservable profitability differences —for example, in the case where minority agents are relatively “unprofitable”— may result in finding no in-group preferences within minority groups even when one exists, or an in-group preference among majority groups even when none exists. Finally, it is difficult to assess the efficiency of outcomes in most economic trans- actions —the sale price of an automobile (as in Ayres and Siegelman, 1995), for example, largely involves the distribution of a fixed pie.

1 We use data from a large state-owned bank in India that provides a near-ideal setting for studying the extent and rationales for preferential in-group treatment in individual interactions with private information. Specifically, the setting makes it possible to ad- dress the three identification problems highlighted above. Detailed credit and personnel records allow us to match all borrowers and branch head officers to their religion and caste, providing a dyadic characterization of the cultural “distance” between transact- ing parties in the allocation of personal loans for close to three million borrowers over a five year period. An explicit officer rotation policy among branches provides variation in the matching between lenders and borrowers. We are thus able to control effectively for time-invariant attributes of borrowers and lenders, and for time varying credit conditions within narrowly defined geographic markets. Further, using detailed loan records, we can measure the effect of cultural proximity on ex ante loan contracting characteristics and ex post loan performance, and hence characterize the efficiency of transactions as well as the mechanisms driving the differences between in-group versus out-group interactions. In addition to the econometric advantages of our setting, the welfare implications of the relationship between cultural proximity and credit outcomes are likely to be of first order in an environment characterized by credit rationing and a long history of religious and caste conflict.1 We find strong evidence of preferential in-group treatment. In the baseline results we define two individuals to belong to the same group when both are members of the same minority religion (Christian; Muslim; Sikh; Parsi; Buddhist) or, conditional on belonging to the majority religion (Hindu), when both belong to the same official caste (General caste, Scheduled Caste, Scheduled Tribe, or Other Backward Classes). On average, the total amount of credit outstanding to borrowers in a group increases by 18.6% when the officer assigned to the branch belongs to the same group. Having an in-group officer also increases the number of borrowers by 6.2% and the probability that the group receives any credit by 1.6%. The results are robust to the inclusion of district-group-time dummies, and

1For evidence and discussions, see Banerjee, Cole, and Duflo 2004, Banerjee and Duflo 2008, and Field et al. 2008.

2 also the simultaneous inclusion of branch-time and group-time dummies. This indicates that the estimated effects are not driven by unobserved variation in the demand for credit by any group or at any locality, by policies that direct credit differentially to different groups and regions over time, or by reverse causality, where officers are transferred to areas where her group is thriving. The results are also robust to an alternate and independent classification where we use individuals’ surnames to assign borrowers and officers based on the religious caste system that prevailed in ancient India. This rules out the possibility that the results are driven by systematic errors in the classification of individuals in the bank’s records. Having established that borrower-officer cultural proximity has a causal effect on ac- cess to credit and the amount provided, we explore the economic mechanism behind this result by analyzing the effect of proximity on loan repayment and other dimensions of credit supply. Loans made to in-group borrowers have better repayment performance ex post. The economic magnitude of this effect is large: weighted by loan size, in-group borrowers are nearly 15% less likely to be late in loan payments, an effect that persists even after the in-group officer is replaced by an out-group one. This improvement in credit risk is recognized ex ante, as in-group loans are made with lower collateral ratios. Additionally, we observe that cultural proximity increases substantially the dispersion of lending across borrowers, implying that officers increase credit to some borrowers more than others within their own groups. These findings suggest that cultural proximity miti- gates problems of asymmetric information in lending and improves the allocation of credit across borrowers with heterogeneous repayment prospects. These effects dominate any negative impact of taste-based preferences on loan quality, as taste-based discrimination would lead —at least weakly— to a decline in average repayment performance. Our results all hold for the subsample of borrowers that had already had loans from the bank when a new officer arrives. Further, our results hold even in isolated locations with a relative scarcity of other formal lenders. Thus, the patterns we report cannot be the result of self-selection by borrowers or the switching of borrowers from other banks.

3 This has the further implication that the observed in-group lending increase improves access to credit or substitutes for informal sources of credit, which are typically more expensive.2 This implies that cultural proximity not only improves the efficiency of the bank’s credit allocation, but it also increases borrower welfare. Our work has a number of significant economic and policy implications that relate to several areas of research. First, our findings highlight the fact that the information and enforcement advantages of cultural proximity can be mistaken for discrimination. For example, in the context of our paper, Hindu borrowers represent 89.2% of the borrower population and take out larger loans than minority religion borrowers —presumably be- cause they are wealthier and in other ways better credit risks than others— yet Hindu officers do not exhibit in-group preferential treatment based on religion. A naive regres- sion of loan access on borrower religion would indicate discrimination against minorities rather than preferential in-group treatment among minorities. This calls for caution in the interpretation of, and policy prescriptions that can be derived from, a substantial body of research devoted to studying minority discrimination that identifies differential treatment based solely on the identity of one of the parties of the transaction.3 Our findings also suggests a mechanism for the formation of statistical discrimination against a minority group, arising as a consequence of the documented effect of cultural proximity on credit outcomes. Even if all groups in the population are ex ante equal, our results imply that the average minority group borrower will have a worse credit history and lower access to credit only because of the low probability that she will be matched with an officer from her own group. If lenders’ priors on borrower creditworthiness are based on the group’s (unconditional) average past performance, minorities will face higher borrowing costs in the marketplace due purely to statistical discrimination. This insight

2See Aleem 1990 and Banerjee and Duflo 2010. 3See, for example, Goldin and Rouse 2000, Bertrand and Mullainathan 2004 and Charles and Guryan 2008 for evidence in labor markets, and List 2004 for evidence in sportscard trading markets. There is also evidence of discrimination for different types of credit markets, such as mortgages (see Ross et al. 2008 for one recent example, and Ladd 1998 for a survey of the evidence), small business lending (Blanchflower et al. 2003), trade credit (Fafchamps 2000; Fisman 2003), and online person-to-person lending (Pope and Sydnor 2010).

4 relates to a body of theoretical work, following Arrow (1973), that rationalizes statistical discrimination as an equilibrium with self-confirming beliefs, but that is silent about the origin of these beliefs. A policy that increases the likelihood of a group match between lenders and borrowers would unambiguously improve efficiency when statistical discrimi- nation is a consequence of in-group preferential treatment.4 We contribute to two main empirical literatures. First, our work relates to the set of studies that examine the role of group identity using dyadic data, with generally mixed results. Ayres and Siegelman (1995) finds evidence of race and gender discrimination in an audit study of price bargaining in the U.S. new car market, but finds no evidence of in-group preferential treatment. In contrast, Parsons et al. (2011) documents that Major League Baseball strikes are called less often if the umpire and pitcher do not match race or ethnicity, consistent with preference-based discrimination. Schoar, Iyer and Kumar (2008) present evidence in a setting similar to ours: bargained prices are lower when the buyer and seller belong to the same community in a field study in the wholesale market for pens in India. The key contribution of our study is to provide the first characterization of potentially efficiency-enhancing mechanisms behind in-group preferential treatment. Our paper also relates to the literature on social ties between transacting parties. Be- cause transacting parties in our setting share a common cultural endowment, but have most likely never met before the transaction, the impact of cultural proximity that we document here is distinct from the effects of social ties or networks, which result from parties’ past interactions.5 Our results indicate that cultural endowments affect the like- lihood that two individuals will interact. This suggests that the endowment and social ties effects are typically confounded in existing work that associates endogenous past so- cial interactions with future market transactions.6 The distinction is important because

4See Kim and Loury 2009 for a discussion of the origin of statistical discrimination, and Coate and Loury 1993, Norman 2003 and Fryer and Loury 2005, for discussions of optimal policy prescriptions in such multiple equilibrium settings 5For evidence of the effect of social connections on economic interactions see, for example, Banerjee and Munshi 2004 and Bandiera, Barankay and Rasul 2009. 6For examples of this work see Cohen, Frazzini and Malloy 2008, 2010, Hwang and Kim 2009, Engel- berg, Gao and Parsons 2011, Jackson and Schneider 2011, and Li 2012.

5 cultural endowments, such as religion and caste, are assigned at birth and transmitted across generations of individuals of the same group, while social ties and connections are dynamic and often subject to individual choice (Becker 1996). This implies that the eco- nomic consequences of cultural endowment differences across groups can persist in the long run, and potentially perpetuate inequality. In the next section, we begin by providing an overview of the data and a description of the Indian bank we study —its organization, the incentives of its officers, and so forth. In Section 3 we present the baseline empirical specification for the analysis. Section 4 presents our results on lending quantity; Section 5 analyzes default, loan dispersion, and collateral patterns to distinguish between taste versus information and enforcement based explanations; Section 6 presents our findings on the cross-sectional heterogeneity in the effects of cultural proximity. In Section 7 we conclude with some policy implications and directions for future work.

2 Data

The main variables in the analysis are obtained from the individual loan portfolio and personnel records of a large state-owned Indian bank, which operates over 2000 geograph- ically dispersed branches in India (see Figure 1). The sample starts in 1999 Q2 and ends in 2005 Q1. This section describes in detail the structure and construction of the dataset from the bank as well as other sources, and relevant background information on the organization of the bank itself.

2.1 Loans, Borrowers, and Branch Heads

The individual loan portfolio data include loan-level information for every borrower with a loan outstanding during the sample period (2.92 million individuals), with information about the loan contracts and repayment status reported on a quarterly basis (1.23 million borrowers per quarter on average). The main variables for the analysis are the amount of

6 debt outstanding, the collateral posted, and the number of days late in interest payment. The median (mean) amount of debt outstanding in the full borrower-quarter panel is 8,495 (36,086) rupees; excluding borrower-quarter observations with zero balance, the median (mean) is 14,645 (47,924). Outstanding debt is typically secured: the median collateral to loan ratio in the full panel is 1.67. The over-collateralizing reflects that fact that the collateral data in most cases are not updated as a loan is repaid, and reflect the initial collateral value at the time of contracting. The median interest rate is 12%. The median borrower’s interest payments are current. The skewness in days late results from the stock of past defaulted loans, even those that occurred prior to our sample period, which are never removed from bank records. Excluding observations with more than 365 days late in repayment (2.76 million borrower-month observations), the average days late is 13.4. To eliminate the noise generated by these legacy of defaults, we retain in the sample only borrowers with less than 365 days late in repayment for the purposes of the analysis.7 From the internal personnel records of the bank, we obtain information about employ- ees at each branch at a quarterly frequency. Each record has a general job description and the position in the internal hierarchy of the bank. We use these data to identify the head officer of each branch at each point in time (4,270 distinct officers in total). Loan officers are classified into six grades, increasing in seniority, and the ability to approve larger loan amounts. The highest ranked officer in each branch is the branch head. For smaller branches, the head officer may himself have a relatively low grade. This implies that any larger loan request that comes through the branch will have to be approved by a higher grade officer elsewhere in the region. Still, in these cases the decision of whether to submit the loan for approval at a higher level of the bank hierarchy is under the head officer’s discretion, and based on information collected at the branch level. Although of- ficers have control over loan and collateral amounts, they have no discretion over interest rates, which are determined by headquarters based on loan type. For example, all home

7The conclusions are invariant to the cutoff of days late used to generate our main sample, although the noise introduced by including the legacy loans generates attenuation bias.

7 improvement loans pay the same rate, as do all educational loans above Rs.400,000. Branch heads —the focus of our analysis here— are evaluated annually on a range of criteria.8 These include quantitative measures such as the amount and profitability of lending, as well as qualitative considerations such as employee skill development, effective customer communication, and other aspects of “leadership competency.” Each officer is ultimately assigned a numerical grade from zero to one hundred. One specific aspect of officer performance that will be relevant for our analysis is the extent to which officers are held accountable for loan defaults after moving branches. Typically, officers are re- sponsible for loans they approve for three years following their departure, at which point responsibility is transferred to an officer in the branch where the loan was made. While there is limited incentive pay, branch heads may be motivated through possible promotion to higher grades or better postings. For example, successful branch heads may be sent to locales with more or better perquisites, such as higher pay (overseas), larger houses, the use of a car, or control over a larger portfolio (large branches). In the analysis that follows we evaluate the extent to which such endogenous allocation of officers to branches affects our estimates.

2.2 Religion, Official Caste, and Religious Caste

The bank data also contain information on the religion and official caste classifications of each borrower and employee. Individuals are grouped into seven categories based on the prominent religions in India: Hindu, Muslim, Christian, Sikh, Parsi, Buddhist and others. They are also classified into four castes based on the categories explicitly recognized by the Constitution of India: General Class (GC), Scheduled Castes (SC), Scheduled Tribes (ST) and Other Backward Classes (OBC). The SC category comprises all the castes historically treated as “untouchable” by the upper castes in India. The ST category includes indigenous, typically geographically isolated, tribal groups. The

8Information on evaluation and compensation of managers within the bank come primarily from interviews with bank staff.

8 OBC category is a collection of caste groups ranked above untouchables in the ritual hierarchy, but socially and educationally disadvantaged. Individuals belonging to the SC, ST, and OBC categories receive targeted government aid and benefit from positive discrimination policies (subject to means testing) such as reservations in public sector employment and higher education.9 Although the SC, ST, and OBC categories include a wide variety of social groups across India, locally they are often relatively homogeneous. The GC category is essentially a collection of all the individuals not belonging to the aforementioned “backward” classes. In order to obtain a group classification that is independent of the bank’s records, we use the borrower and officer surnames to generate a group classification based on religious castes. According to religious texts such as Manusmriti, Hindu society is broadly divided into four Varnas: the Brahmins (priests and scholars), Kshatriyas (warriors), Vaishyas (merchants and traders) and Shudras (laborers and artisans). Each Varna is a unification of several Jatis, or communities (see Buhler 1886), and a person’s surname typically reflects the Jati she belongs to. We exploit this link with surnames to classify each individual into her Varna (see Banerjee et al. 2009 for a further discussion of the link between surnames and castes in India).10 The following are examples of the name matching and search process using three common surnames in India: 9The categories of Scheduled Caste (SC) and Scheduled Tribes (ST) that represented a majority of lower-status castes as well as tribes were first protected in anti-discrimination laws through the ninth schedule of the Constitution in 1950 (Article 15, 17, and 46). In 1990, the further caste-based catego- rization of OBC was added for identifying additional socially and economically deprived communities. A few years thereafter the category of OBC was extended to include a significant segment of the non-Hindu population, notably Muslims, Christians, and Sikhs. 10Since the association between individual names and their borrowing and employment records is proprietary and cannot be disclosed outside the bank, the process of assigning individuals to the Brah- mins, Kshatriya, and Vaishya groups followed four steps. First, the bank provided us with a list of all surnames—both borrowers and officers—present in bank records. Second, we searched Google and the Anthropological Survey of India (Singh, et al., 1998, 2003, 2004) to establish a community association for each name. Third, we searched Google, Wikipedia, matrimonial websites, and other references (Dahiya 1980, Dudhane 2007, UNP, Marathas 2010, Maheshwari Samaj 2006, Bindu 2008) to establish the link between communities and Varnas. Finally, the bank linked community and Varna information to bank records by surname, before removing the borrower and manager identifiers from the data.

9 • Example 1. Surname: Birla. A Google search of the surname was found listed in one of the matrimonial sites of the Maheshwari Samaj community (Maheshwari Samaj 2006). In the Maheshwari Samaj we find information that Birlas belong to the Vaishya Varna.

• Example 2. Surname: Rathod. The name Rathod was found in the Anthropological survey of India to be commonly used by the Rajput community (K. S. Singh et al., 2004). Following up with K. S. Singh et al. (2004) we find that the Rajputs are Kshatriyas according to the Varna system.

• Example 3. Surname: Deshpande. A Google search showed the surname listed under the Deshastha community.11 A search in Kamat.com showed this community as belonging to the Brahmin Varna.

Classifying borrowers and officers by religious caste using surnames comes at the cost of additional noise and imprecision. One source of noise results from the fact that many surnames can be classified into two or more Varnas. For example Saxena is grouped under both Brahmins and Kshatriyas. Similarly Desai is grouped under both Brahmins and Vaishyas. We created three special categories for individuals where this ambiguity arises (Kshatriya-Brahmin, Kshatriya-Brahmin-Vaishya, and Kshatriyas-Vaishyas). We note, however, that once we condition on region — as we do throughout our analysis — there is a clearer link between names and communities. Second, it was unclear how to categorize individuals into the Shudra Varna according to their community affiliations, which precluded using surnames for individuals outside of the General Classes. Finally, in a large fraction of cases, the surname-based classification conflicted with the bank classifications assigned to loan officers and borrowers. For example, many bank-classified Muslims had “Hindu” surnames, and vice-versa. Still, exploring the effect of proximity along the Varna dimension is interesting it its own right, and will allow us to ascertain whether the results based on bank classifications are driven by systematic misclassification of officers and borrowers in the bank records. 11http://en.wikipedia.org/wiki/List of Deshastha Brahmin surnames.

10 2.3 Descriptive Statistics - Group Composition

The religion, official caste, and Varna compositions of the borrowers and officers popu- lations are shown in Table 1. By religion, Hindus represent the majority of borrowers (89.4%) and officers (93.8%). The largest group of minority borrowers is Muslim (6.33%), and the largest officer minority is Christian (2.1%). Hindus are over-represented and Muslims under-represented in the borrower and officer populations relative to the total population (80.5% Hindu and 13.4% Muslim according to the 2001 census). Most bor- rowers and officers do not receive any official designation and are classified as General Class (66.7% and 74.3% respectively). The largest borrower minority is the OBC cate- gory (16.6%), while the largest officer minority is ST (15.7%). SCs are under-represented in the borrower sample and STs under-represented in the officer sample, relative to the population (16.2% SC and 8.2% ST in the 2001 census).12 We are able to match surnames to Varnas for a subsample of the population; a total of 502,723 borrowers (18.3% Brahmin, 60.5% Kshatriya, 6.6% Vaishya, 1.7% mixed cate- gories, and 5.72% in other categories) and 1,689 officers (23.0% Brahmin, 43.4% Kshatriya, 11.7% Vaishya, 15.5% mixed categories, and 6.4% in other categories) have Varna assign- ments. All the identifiable Varnas in our sample belong to the General Class according to official caste definitions.

2.4 Descriptive Statistics - Branches and Groups

Table 2, Panel 1, shows branch-quarter level statistics. The median branch has a total of 13.5 million rupees of debt outstanding and lends to 334 borrowers. The borrower composition is generally quite heterogeneous: the median branch lends to borrowers of four different religions and three different official castes. The median branch is small, with two loan officers including the head officer, and the modal branch has only a single officer. 12The 2001 India Census does not keep track of OBCs.

11 The unit of analysis is the branch-group-quarter level (indexed by b, g, and q, re- spectively) where group refers to the cultural group of the borrower.13 In our main specification we use the full set of religion and caste information to group borrowers into 9 categories: 5 minority religions, and 4 official castes conditional on religion being Hindu. In other specifications we consider group definitions based on religion, caste, and Varna independently. The panel employing our main group classifications has 339,366 branch- group-quarter observations and we present the descriptive statistics in Table 2, Panel 2. The average group-branch has around 2 million rupees of total debt outstanding and 43 borrowers, with an average of 16 days late in repayment. 8.1% of borrowers are more than 60 days late in their loan repayments. However, only 2% of total debt is in default, reflecting the fact that larger debts are much less likely to be behind in repayment. We merge the branch-level personnel information to this panel to obtain our main explanatory variable, SAMEGROUP bgq, a dummy variable that is equal to one for the branch-group-quarter loan cells where the branch head officer belongs to group g, and zero otherwise. For example, if the head officer of branch b in quarter q is Muslim, then

SAMEGROUP bgq = 1 for loans from group g if g = Muslim, and zero for the other groups in that branch-quarter.

2.5 Officer Rotation

The bank follows an explicit policy of geographical officer rotation, with the stated ob- jective of reducing opportunities for corruption, nepotism, and other perverse incentives in the allocation of loans. As a result, branch turnover is high: we observe an average of 127 head officer reallocations per quarter, and the median branch has one officer change during our sample period. The mean (median) spell of a head officer in a branch is 8.3 (8) quarters, with standard deviation of 4.1. Head officers are always assigned to branches

13We do not perform our analysis at the level of the individual borrower for two reasons. First, the relevant individual-level decisions in our context are at the head officer level, which is the level of aggregation we use. More practically, it would be very difficult —in some cases impossible, as when we study the effect of proximity on group lending variance— to implement some of our analyses using individual-level data.

12 that are located away from their home town, and the average officer reallocation assigns the officer to a new branch that is 250 kilometers from the previous assignment. This im- plies that although officers generally stay within the same region, it is unlikely that they have had any prior interaction with any of the potential borrowers in their new location. In Appendix Table A.1 we report the empirical distribution of branch transitions by religion (panel 1), official caste (panel 2), and Varna (panel 3). We highlight in boldface the transition frequencies that are different from those that would result from assigning officers at random from the population of officers described in Table 1. For religion and Varna, the empirical transition frequencies are statistically indistinguishable from a policy of random officer allocation. However, for the subsample of Hindu officers the empirical transition rates deviate from the random benchmark when officers are grouped by official caste. For example, the observed probability of a GC to GC transition is 61.0%, while the random benchmark is 55.4%, and the difference is statistically significant at the 1% confidence level. This indicates that there are some branches that tend to receive General Class officers too often relative to random assignment. This is consistent with the existence of reservations for SC, ST, and OBC official posts in some regions, since this would lead to a higher observed proportion of transitions in the diagonal of the matrix. We discuss the potential consequences of policy-driven distortions in officer allocation in the context of our empirical estimation in the sections that follow.

3 Empirical Specification

Our baseline empirical specification takes the following form:

ygbq = βSameGroupbgq + αgb + τq + bgq (1)

The dependent variable is a loan outcome (i.e., total lending, number of loans, fraction of loans past 60 days late, etc.) at the branch-group-quarter level; g indexes the group

13 —caste, religion, or pooled partition; b indexes the branch; and q indexes the quarter. SameGroup is an indicator variable denoting whether the branch head in branch b belongs to group g in quarter q. The two fixed effects —agb and τq— capture time-invariant attributes of each group within each branch (i.e., a group times branch set of fixed effects), and aggregate shocks to all branches. The error term bgq allows for clustering at the branch level. We also consider augmented specifications where we saturate the branch-time, group-

Branch time, state-group-time, and district-group-time variation. Branch-time dummies, τbq , account for all changes in the demand for credit in a particular location, as well as changes in directed credit policies aimed at certain localities. Since there is one head loan officer per branch at any time, the branch-time dummies also account for all unobserved branch head heterogeneity, whether time invariant (different officer skills) or time varying (officer learning about the job or the environment), and also for any effect that the change of an

Group officer may have on average lending in a branch. The group-time dummies, τgq , account for aggregate changes in the credit demand from, or supply to, specific groups. The state- State-Group District-Group group-time (district-group-time) dummies, δbgq (δbgq ), absorb any changes in the demand or supply of credit that are specific to a group in any given region, such as secular borrowing trends affecting particular groups in a location. We show in the next section that these saturated models produce estimates of the coefficient on SameGroup, β, that are statistically indistinguishable from those of the baseline. This represents strong evidence that officer group transitions are orthogonal to other determinants of credit at the group-location-quarter level (i.e., E[SameGroup.] = 0), and that the β estimates from specification (1) are consistent. The coefficient on SameGroup is a difference-in-differences estimate of the effect of cultural proximity between a lender and a borrower on loan outcomes. Consider, for example, the regression with the log of total lending as the dependent variable and, for simplicity, suppose there are only three groups: Hindus, Muslims, and Christians. Suppose that a branch has a Hindu officer during the first half of the sample, and a Muslim

14 officer during the second half. The coefficient on SameGroup captures the difference between the log debt of Hindu borrowers in the branch when the officer is a Hindu (in- group) officer relative to when the officer is a Muslim. It also captures the difference between the log debt of Muslim borrowers with a Muslim officer relative to a Hindu one. Our main results show the average effect across all groups. In the last section of the paper we explore whether there are differential in-group effects across group definitions. Specifically, we include the interaction of SameGroup and a dummy for each group g. We also explore the heterogeneity of the effect along other branch and branch location characteristics.

4 Results: Loan Amounts

We begin with a graphical description of (unconditional) lending patterns around officer transitions. First we classify borrowers into two categories based on whether they have the same group identity as the outgoing officer: in-group borrowers are those belonging to the same group as the officer, and all others are categorized as out-group borrowers. For example, in a branch where the outgoing officer is Hindu, the Hindu borrowers are in-group before the officer change, and all minority religion borrowers are classified as out-group. Each of these borrower groups may or may not experience a change in their in-group/out- group status after the officer change. For example, suppose the Hindu officer is replaced by a Muslim one. Then, Hindu borrowers transition from in-group to out-group, Muslim borrowers transition from out-group to in-group, and other non-Muslim minority religions remain as out-group. Alternatively, if the replacement officer is also Hindu, then Hindu borrowers remain as in-group and all minority borrowers remain as out-group. We use these borrower classifications to construct “event study” plots around officer transitions. The horizontal axis of the plots in Figure 2 measures time in quarters since the officer change in a branch. Time 0 represents the first quarter a new officer appears as the branch head in the personnel files. Given that our analysis is based on quarterly

15 data, the new officer may arrive up to 11 weeks before the observed entry time. Below, we discuss the bias that this measurement error may introduce. The vertical axis measures the average debt of borrowers that experience a change in in-group/out-group status, relative to those that do not. Panel (a) shows the average debt of out-group borrowers that become in-group after the officer change, minus the average debt of borrowers that remain out-group. Panel (b) shows the average debt of in-group borrowers who become out-group, relative to in-group borrowers that remain in-group. All averages are taken at the group-branch level using the caste-religion group definition, and both plots include the 95% confidence interval of the mean difference. The plots show a shift in the composition of lending as a function of cultural ties when there is a change in head officer. In Panel (a), the average debt of borrowers that switch from out-group to in-group status increases by approximately 1.5 million rupees (approximately USD$35,000), relative to borrowers that remain out-group after the officer change. A parallel pattern appears in Panel (b): the average debt of borrowers that switch from in-group to out-group status drops by 4 million rupees during the four quarters following the change in status, relative to borrowers that remain in-group.14 The plots in Figure 1 also suggest that the relationship between cultural proximity and lending may be causal, since the relative debt change occurs immediately around the officer transition and does not appear to be driven by pre-existing differential lending trends across the two groups in each panel. The plots provide some validation for the identification assumptions behind the difference-in-difference estimator of the in-group effect in specification (1). There is a small and statistically insignificant increase in the relative amount of lending at time -1 in Panel (a), the period prior to the recorded quarter of arrival of the new officer. This is likely driven by the aforementioned measurement error in the time of arrival of the new officer. Such measurement error will tend to bias towards zero our estimates of 14The asymmetry in the magnitudes of the jumps across the two plots is driven by the fact that the groups of borrowers compared in the two plots are different in size and average debt. In the formal empirical analysis these differences are accounted for by the branch-group dummies.

16 the in-group effect in specification (1).

4.1 In-Group Effect on Credit

We present in Table 3 the effect of having an in-group branch head on credit, estimated using specification (1). Outcomes are measured at the level of group g in branch b in quarter q. To capture both the intensive and extensive margins of lending, we use as dependent variables the log of total debt, the log of the number of borrowers, and a dummy variable equal to one if debt is greater than zero. The log transformation of the first two variables reduces their skewness, and facilitates an elasticity interpretation of the coefficients. However, it also creates an unbalanced panel, as zero loan cells become missing values. The effect on the group-level extensive margin —i.e. the probability that a group receives some credit— is captured by the last specification in a linear probability model. In unreported specifications we used a ln(1+x) transformation to reduce the skew- ness of the dependent variable. Although this alternate transformation does not allow a ready interpretation of the magnitudes of estimated coefficients, it generates qualitatively identical results to those reported below. The estimated coefficients on the SameGroup indicator variable are positive and sig- nificant across all three specifications. The magnitudes imply that lending to a group increases by 18.6% when an in-group officer is assigned to the branch (Table 3, column 1), the number of borrowers increases by 6.2% (3, column 4), and the probability that the group receives any debt increases by 1.5 percentage points (3, column 7), or around 2.5% of the baseline probability of having positive credit. Because cultural proximity affects the likelihood that a group receives any credit, the coefficient estimates on total lending and number of borrowers are consistent estimators for a local average treatment effect (LATE) of cultural proximity on branch-group cells that have positive debt both with and without an in-group officer. The LATE and the average treatment effect are likely to be close in our setting, because the effect of proximity on a group’s probability of borrowing is low. To confirm this we estimate the bounds of

17 the average treatment effect using the procedure in Lee (2008), estimating the total debt and number of borrower specifications after trimming the 1.5-percent upper and lower tails of the outcome for the branch-group-quarter cells in which SameGroup equals one (1.5 percent is the estimated treatment effect on the probability of receiving any credit). In both cases, the bounds are relatively tight around the LATE estimate, as the difference between the LATE and the lower bound is less than one fifth of a standard deviation of the LATE estimate (Table 3, columns 2, 3, 5, and 6). For example, the average treatment effect of cultural proximity on total debt ranges from 0.183 to 0.229, while the LATE estimate is 0.186. Thus, in what follows, we present LATE estimates and interpret them as the lower bound on the average treatment effect of cultural proximity on loan outcomes. Table 4 presents the estimates of specification (1) augmented with branch-quarter, group-quarter, state-group-quarter, and district-group-quarter dummies.15 To reduce the number of nuisance parameters to be estimated in these specifications, we remove the branch-group means from all variables in the panel rather than including branch-group fixed effects, and adjust the standard error estimation accordingly. Neither the magni- tude nor the significance of the estimated parameters changes in any of the saturated models. This represents strong evidence that our variable of interest, SameGroup, is un- correlated with the error term in the baseline specification (1), and that the estimated coefficient on SameGroup has a causal interpretation. The saturated specifications rule out, for example, an alternative reverse causality interpretation of the coefficient in which a positive association between lending to a group and the identity of the officer is driven by an endogenous allocation of officers into areas where their own group is thriving. The district-group-quarter dummies account for localized shocks and trends to the demand for credit of specific groups in narrowly defined geographical areas (the median district has three bank branches in our sample). Table 5 repeats the estimation using the group definitions based on the traditional religious caste system (Varna), obtained through surname matching. The estimated ef-

15There are 43,974 branch-quarter dummies, 216 group-quarter dummies, 4,760 state-group-quarter dummies, and 52,406 district-group-quarter dummies

18 fect of cultural proximity on lending is again positive across all outcomes, although the estimates are noisier (e.g. the effect on the number of borrowers is not significant). The point estimates are of the same order of magnitude as those obtained using our main group definitions based on religion and government-sanctioned caste (Table 3). The point estimates imply, for example, an in-group effect on total credit of 14% when groups are de- fined using Varnas (18.6% when defined using religion and government sanctioned caste). The Varna grouping is constructed independently of the bank’s classification of officers and borrowers, indicating that the observed in-group effects are not driven by systematic misclassification of borrowers. Also, since it is implausible (and illegal) that the bank uses Varnas to allocate credit or assign jobs, the Varna-based results provide an indepen- dent validation of the identification assumption that the group identity of the officer in a branch is uncorrelated with directed lending policies targeted to borrowers of the same group.

4.2 Intensive and Extensive Margins

The results so far focus on lending outcomes at the group level. In Table 6 we explore the effect of cultural proximity on the borrower-level intensive and extensive margins of lending. To do this we partition the borrower sample into two groups: 1) borrowers that established a credit relationship with the bank prior to the arrival of the current officer, and 2) borrowers that receive credit from the bank for the first time with the current officer. Panel 1 shows the estimates on the first subsample, which includes every branch- group-quarter bgq where there exist borrowers that had positive credit at any time before the officer in charge of branch b in quarter q arrived. The estimates indicate that the average debt by existing borrowers increases by 11.6% with the arrival of an in-group branch head. The coefficient on the number of borrowers in this subsample measures exit, because all entrants are removed by construction. The effect on the number of borrowers is a tightly estimated zero, indicating that cultural proximity has no effect on the probability

19 that preexisting borrowers cease to borrow from the bank. These results imply that the estimated effect of cultural proximity on the total number of borrowers from Table 3 is solely due to its effect on entry. Collectively, the results on the existing borrower subsample imply that the effect of proximity on credit is positive holding the borrower pool constant. The estimated effect on the probability that a group receives any credit is positive, indicating that proximity increases the likelihood that borrowers that had borrowed and repaid their loans start to borrow again. Table 6, Panel 2, shows the estimated effects for the complementary subsample of borrowers that obtain credit from the bank for the first time with the current officer. The estimated effects indicate that cultural proximity increases the flow of total lending to new borrowers by 16.9% and the number of new borrowers by 13.3%; the difference between the two estimates indicates that the average size of in-group loans is larger. The effect of cultural proximity on entry can occur through two different, not mu- tually exclusive, channels. First, cultural proximity may increase the likelihood that an officer approves a loan application from an in-group borrower. Second, it may affect the likelihood that in-group borrowers apply: borrowers with rational expectations about the higher likelihood of having an application approved by an in-group officer are more likely to request a loan. For example, a Muslim borrower may decide to save on the cost of the loan application if she knows that the branch head officer is non-Muslim, but choose to apply once a Muslim officer is in place. The data do not contain information on loan applications, so we cannot estimate the effect of officer identity on applications directly. We adopt an indirect approach and analyze the dynamics of the in-group effect on the intensive and extensive margins. If entry increases because officers approve more in-group applications, then the in-group effect on the number of borrowers should occur immediately after the officer’s arrival. In contrast, the effect of an in-group officer on the inflow of new applications should be more gradual, as it takes time for the news that there is an in-group officer in the branch to

20 spread. We estimate specification (1) augmented with interactions between SameGroup and a set of indicator variables for the officer’s first quarter at the branch, his second quarter, and so on. The coefficients on the interaction terms represent how the effect of cultural proximity on loan outcomes changes during his tenure. Figure 3 plots the estimated interaction coefficients and 95% confidence interval bounds using the log of total credit and number of loans as dependent variables. Given that cul- tural proximity affects only the entry margin, we present estimates on the full sample and interpret the effect on the number of loans as the effect on entry. Panel (a) shows that the effect of cultural proximity on the amount of credit is immediate: credit to in-group borrowers increases by close to 10% during the first quarter the officer is assigned as branch head. In contrast, the effect on the number of borrowers, although positive, is not statistically different from zero when the officer arrives in the branch (Panel (b)). The effect of cultural proximity on borrower entry becomes significant only six months after the officer’s arrival. The distinct dynamic patterns in the two plots suggest that the effect of cultural proximity on the intensive and extensive margins is driven by different economic phe- nomena. For the intensive margin, the officer increases lending immediately to existing in-group borrowers. The effect on the extensive margin takes time to build, and is con- sistent with officers attracting more in-group borrower applications. We discuss further the implications of these patterns after presenting the results on loan performance. Our main conclusion thus far is that cultural proximity between lenders and borrowers leads to an increase in lending. The magnitudes of the estimated effects on access to credit and the amount borrowed are substantial. The next section investigates the potential mechanisms behind this preferential treatment of in-group borrowers.

21 5 Preferences, Enforcement and Screening

The key identification challenge in understanding the causes and consequences of prefer- ential in-group treatment is whether it is due to favoritism or reduced frictions in loan transactions. Distinguishing amongst causes is critical to understanding the efficiency implications of preferential in-group lending. It is possible to narrow down the set of explanations by studying the impact of in-group lending on ex post loan performance and on other dimensions of lending. We consider three main models of preferential in-group treatment: (a) pure favoritism, i.e., taste-based preferences, with no difference in lending frictions; (b) reduced ex ante frictions (i.e., better screening); and (c) reduced ex post frictions (i.e., improved enforce- ment). We will refer to the frictions in (b) and (c) collectively as asymmetric information. There are various mechanisms that may allow in-group officers to mitigate the effects of information asymmetries. For example, loan officers may be better able to enforce repay- ment by borrowers that belong to her own group because it lowers the cost of imposing penalties thorough social sanctions. In a similar vein, social norms or altruism may in- crease borrowers’ utility cost of defaulting when the issuing officer belongs to their own group. Cultural proximity may also reduce ex ante information frictions by, for example, providing the screening officer a better signal of a borrower’s creditworthiness. In general, if loan officers can mitigate the effects of asymmetric information for in- group borrowers, we expect the expansion of in-group credit access – already documented above – to be accompanied by improved repayment. By contrast, if favoritism is the dominant source of within-group preferences, the increase in lending will be the result of credit expansion to (lower-quality) marginal borrowers, leading to a deterioration in average lending quality. We may distinguish between the most straightforward models of ex ante and ex post information frictions. If enforcement were the dominant mechanism for improved re- payment performance, we would expect no change in the distribution of loan size, since enforcement primarily affects ex post outcomes. Further, enforcement takes place through

22 sanctions imposed by the loan officer himself, we expect the performance of in-group loans to decline after an in-group officer’s departure. We find that in-group loans have higher repayment performance, evidence that fa- vors screening and enforcement models, and implying that on balance in-group lending reduces information frictions. Further, we find that the effect of cultural proximity on loan repayment persists even after an in-group branch head is replaced by an outgroup one, and that loan size dispersion is higher for in-group loans. These findings suggest that the higher repayment performance of in-group loans can be attributed at least in part to better in-group screening.

5.1 In-Group Effect on Loan Performance

We examine the impact of cultural proximity on loan performance by estimating specifi- cation (1) using the fraction of borrowers and debt that are more than 60 days past due; the results are almost identical when we use 30 and 90 days past due (see Appendix Table A.2).16 As before, the unit of analysis is the branch-group-quarter level, and our outcomes of interest are calculated over all loans that are active and due in branch b, group g, quar- ter q. We calculate the fraction of loans past 60 days overdue, which weights borrowers equally, and also the fraction of debt overdue, which is a loan-size weighted statistic. In some specifications we look at the 12-month forward default figures, that is, the fraction of loans more than 60 days overdue in quarter q + 3 of all loans active and due in quarter q. The estimated coefficients on SameGroup are presented in Table 7. The point es- timates of the effect of cultural proximity on the fraction of loans more than 60 days overdue —contemporaneous and 12 months forward— are negative and significant at the 1% level (columns 1 and 2). For the contemporaneous default specification the coefficient

16We also employed specifications that used the log of one plus the number of days late as the outcome variable, including both specifications that weighted days late by loan size and specifications that weighted all borrowers equally. These regressions generated qualitatively very similar results to those we report in the text, but do not have any clear economic interpretation.

23 of -0.0057 implies a 7% reduction in default relative to the mean of 8.1% for in-group loans. The effect on default 12 months forward, -0.0097, implies an even larger effect of cultural proximity on loan performance. We obtain similar results when fraction of loans more than 60 days overdue is weighted by loan size, both for the specification using con- temporaneous repayment status and also repayment status 12 months forward (columns (4) and (5)). The coefficient in column (4) is -0.0029, implying a 14.5% reduction in default relative to the mean debt-weighted default rate of 2.0% for in-group loans. Again, the 12 months forward specification is somewhat larger in magnitude, with a coefficient of -0.0054. A taste-based model of higher in-group lending that would also lead to higher repay- ment rates is one where cultural proximity causes loan officers to extend additional loans to insolvent in-group borrowers to make payments on past loans. This “ever-greening” explanation also implies that the impact on loan performance should be relatively short- lived, and in particular that it should disappear when an in-group officer is replaced by an out-group one. We test whether the positive effect of cultural proximity on performance dissipates when the in-group officer is replaced with an out-group one by augmenting our specifications in columns (3) and (6) with the 12 month lead of SameGroup. The coef- ficient on this variable represents the difference in future default across borrowers that still have an officer from the same group relative to those that experienced a change. In (3), which weights all borrowers equally, the coefficient on SameGroup remains large in magnitude and significant at the 1% percent level. The coefficient on the 12 month lead of SameGroup is negative, suggesting an additional further effect of SameGroup in the future; however, it is half the magnitude of the coefficient on the lead of SameGroup and statistically indistinguishable from zero. In column (6), which weights default status by loan size, the coefficient on the 12 month lead of SameGroup is a precisely estimated zero, and the coefficient on the contemporaneous SameGroup is unchanged. Overall, these results argue against an ever-greening explanation for our loan size and repayment per- formance results. Further, the absence of any SameGroup effect in the future evidence

24 that the improvement in loan repayment results from better ex ante screening, as opposed to better ex post monitoring or enforcement, since the cost of direct ex post monitoring or application of social sanctions should increase after the in-group officer’s departure. The patterns in Table 7 are qualitatively similar for the borrower subsamples limited to (a) those who previously borrowed from the bank, and (b) those who borrowed from the bank for the first time in quarter q. The point estimates are uniformly negative for the coefficients of interest, though somewhat larger for the sample of of preexisting borrowers. This suggests that the performance improvement associated with in-group lending is not the result of borrower self-selection. We refer back to our results on the dynamics of the effect of cultural proximity as one additional pattern in the data that helps to rule out borrower self-selection as an explanation for improved loan performance. As we note in the previous section, the increase in lending to borrowers that had already obtained credit from the bank occurs within the same quarter of the in-group officer’s arrival. Combined with the improvement in the performance of these loans, this indicates better screening and monitoring, holding the set of borrowers constant.

5.2 Loan Dispersion and Collateral

The view that in-group lending reduces information frictions and improves allocative efficiency yields further predictions on the ex ante characteristics of in-group lending. First, lower default rates should reduce the average cost of borrowing. Since loan interest rates are fixed in our setting, we focus instead on the collateral as a proxy for the borrowing cost (higher risk borrowers will post more collateral holding the interest rate constant), and examine whether collateral to loan ratios are lower for in-group loans. Second, improved ex ante screening in particular should increase the dispersion in lending, as in the screening discrimination model of Cornell and Welch (1996). The intuition for this prediction is that the precision of the signal that the loan officer receives about a borrower’s quality is more precise for in-group transactions. As a result the

25 variance of the prior distribution of default probabilities is larger. In our setting, this implies a higher dispersion for lending to in-group borrowers. Note that, by contrast, enforcement-based explanations do not make strong predictions about the ex ante loan distribution. Focusing first on effect of cultural proximity on collateral, we employ our baseline specification baseline specification (1), using two separate measures of collateral intensity —the logarithm of total collateral, and the logarithm of the ratio of total collateral to loan amount. The estimated in-group effects are presented in Table 8. Column (1) indicates that the point estimate of the effect of cultural proximity on collateral is 0.136, smaller than the estimated coefficient for total debt (0.185). The estimated in-group effect on collateral to loan ratios is -0.0474, significant at the 1% level, again indicating a lower collateral rate for in-group loans. To assess the effect of cultural proximity on loan dispersion, we estimate the baseline specification (1) in columns (3) and (4) of Table 8, using two measures of loan dispersion within the group: the standard deviation and the interquartile range of the loans issued in branch b, group g, quarter q. The estimated in-group effects, presented in Table 8, are positive and significant for both measures. The point estimates indicate that cultural proximity increases the standard deviation (inter-quartile range) of the loans by 18.3% (8.8%). These findings are most consistent with ex ante information asymmetries accounting for the higher level of lending and performance of in-group borrowers.

5.3 Discussion

The results taken together rule out several standard rationales for the observed effect of cultural proximity on lending. First, the most straightforward taste-based models cannot explain why cultural proximity both increases the supply of credit and improves repayment performance. Preference-based explanations are only consistent with our results if in- group favoritism affects information collection or enforcement. For example, officers may spend more time with borrowers of their own group solely due to preferences and, as

26 a by-product, collect better ex ante information about them. Alternatively, they may feel greater offense by default from in-group borrowers, compelling them to extract higher repayment rates ex post. This class of explanation departs substantially from any standard preference-based discrimination models, since it implies that discrimination may lead to efficiency improvements. Second, simple ex-post enforcement models cannot explain why the repayment per- formance improvement persists after the in-group officers leaves the branch, replaced by an out-group one. Only an enforcement model in which the improvement in repayment does not stem from a direct ex post action by the officer can be consistent with the re- sults, since the marginal cost of such direct actions is very likely to increase substantially with the geographical distance between the officer and the borrower. For example, the borrower may feel guilt or remorse for defaulting on someone from their own community. Given that there is no ex post action required by the lender, this type of model is not different from a standard ex ante screening model in which there are borrower types (e.g. remorseful and immoral), but one where the type of the borrower is contingent on the type of the lender. Third, simple enforcement models with homogeneous agents cannot explain why cul- tural proximity increases the heterogeneity of lending in a group. Continuing with the example above, it would require that there are remorseful and immoral borrowers in every group, but an in-group officer is able to tell them apart. This type of explanation can be easily rationalized with an ex ante screening model where the officers have more precise signals of borrower creditworthiness when the borrower belongs to his own group, as in Cornell and Welch (1996). Overall, the results strongly suggest that cultural proximity reduces asymmetric in- formation problems in credit allocation and may lead to efficiency improvements. We discuss the efficiency consequences further in the next section.

27 6 Efficiency and Heterogeneity

Because the results so far pertain to credit received from only one source, they do not allow for inferences about the welfare and efficiency consequences of cultural proximity. Further, the results on the average effect of proximity may hide substantial heterogeneity that can be informative about the nature of the in-group advantage. In this section we attempt to shed some light on these two issues by characterizing the cross sectional heterogeneity of the effect of cultural proximity along observable dimensions of borrowers, branches, and locations.

6.1 Branch Density

If credit from the bank in our data substitutes rupee-per-rupee for credit obtained from other sources at the same cost, then the in-group effect on credit outcomes does not have any consequences for credit access. This is a priori unlikely because the cost of borrowing from the government bank in our data is subsidized, and is either the lowest or amongst the lowest cost available sources of funding from any formal or informal institution in India. To study formally whether substitution is present, we analyze how the in-group effect varies in the cross-section of districts classified by the density of bank branches, which serves as a measure of the availability of outside borrowing opportunities. Branch density in a district is measured as the total number of branches from all financial institutions per 1,000 inhabitants. The number of branches per district is ob- tained from the website of the Reserve Bank of India and the number of inhabitants per district from the India Census, both from 2001. The average number of branches per 1,000 inhabitants is 0.81 across the 357 districts with a branch from the bank in our data. There is substantial heterogeneity across districts, with 0.18, 0.54, and 1.88 as the 1st, 50th and 99th percentiles respectively. The districts with the highest branch densities typically correspond to urban areas and the lowest densities to rural ones. We estimate specification (1) adding interactions between SameGroup and indicators

28 for whether the branch is located in the districts in the second, third, or fourth quartile of branch density (interactions between the quarter dummies and the district quartiles are also included). Table 9, panel 1, presents the estimated coefficients for different loan outcomes. The estimated coefficient on SameGroup without interactions corresponds to the in-group effect in branches located in districts with the lowest branch density. For total debt, number of borrowers, and debt variance (columns 1, 2 and 3), the estimates are positive and statistically significant, and the magnitudes are close to those obtained on the full sample. The estimated in-group effect on loan performance is also close to that of the full sample. This implies that the documented effect of cultural proximity on credit outcomes is present even in locations where there are few opportunities for outside credit. Across all outcomes, the coefficient estimates for the interaction terms imply a stronger effect of cultural proximity in more isolated areas. This is the opposite of what one would expect if the results were driven largely by substitution of credit from other formal bor- rowing sources. Except for the loan performance specification, however, none of the interaction term coefficients is statistically significant. Thus, the evidence suggests that cultural proximity increases borrowers’ overall access to credit and reduces the cost of bor- rowing, since the alternative sources of credit in isolated areas are informal moneylenders or micro-lenders that charge higher interest rates.

6.2 Hierarchical Distance

Although the modal branch in our sample has one loan officer and the median branch has two (including the head officer), there are branches that are much larger. The bulk of the information collection process is performed by the head officer in smaller branches, while some of the screening and enforcement activities are likely delegated to other officers in larger ones. We would therefore expect to observe proximity to have a smaller effect in larger branches. Table 9, panel 2, illustrates how the effect of cultural proximity varies as a function of

29 the number of officers at a branch. It presents the estimated coefficients of specification (1) augmented with interactions between SameGroup and indicators for whether the branch has two, and more than two officers (and interactions with the quarter dummies). All the results in the previous subsection are confirmed for the modal branch, which has only one officer (coefficient on SameGroup with no interaction). The point estimates indicate that the magnitude of the effect is larger in these branches relative to the average. The point estimates on the interaction terms are quite noisy across all specifications, but broadly point towards less of an effect of cultural proximity on loan size, dispersion, and default in larger branches.

6.3 Homogeneity and Size

In this final subsection we explore the heterogeneity of the effect of cultural proximity across religions, castes, and group sizes. We do this to assess the robustness of the results to different group definitions, and to examine, for a given group definition, how the effects of cultural proximity are affected by the relative prevalence of groups in the population, since an in-group officer’s informational advantage may be relatively small in cases where the in-group represents a high fraction of the total population. The analysis in this section also sheds light on the consequences of using highly aggregated group definitions. We first focus on religion. To do so we build the branch-group-quarter panel by grouping borrowers into Hindus and each of the five minority religions based on the bank’s classifications. The difference between this panel and the one employed to this point in the analysis is that now loan outcomes for Hindu borrowers are grouped in a single cell. With the religion-based group definition, we define a new variable of interest, SameReligion, equal to one for the borrowers of religion g in branch b in quarter q when the head officer in the branch belongs to the same religion g. In Table 10, panel 1, we show the estimated coefficients for SameReligion interacted with indicators for religion equal to Hindu, and religion equal to Minority Religion (we show the heterogeneity within minority religions in the Appendix Table A.3). The point estimates indicate that the in-

30 group effect on all lending outcomes is a tightly estimated zero for Hindus, when caste heterogeneity within Hindus is ignored (odd numbered columns). For minority religions, the in-group effect on all outcomes has the same sign and a considerably larger magnitude than the average effect documented using SameGroup as a covariate. For example, having a minority religion officer in a branch leads to a 43% expansion in lending to same-minority religion borrowers and reduces the fraction of debt more than 60 days overdue by 0.86 percentage points, or 43% relative to the baseline rate of 2%. We explore two potential explanations for this cross sectional variation of the in-group effect. First, the in-group effect may vary with group size: minority religion borrowers represent, by definition, the smallest groups in the population. Second, the in-group effect may vary with group heterogeneity: the Hindu grouping combine borrowers of a wide array of culture, codes, and language. To explore the group size dimension while holding the group definition constant, we augment the religion specification with an additional interaction term using a dummy for whether the state where the branch is located has an above median population of minority religion inhabitants in the 2001 India census (the fraction of minority religion population in the median state is 0.174). The interaction term with the large minority state dummy has the opposite sign to the term with no interaction for all credit outcomes (even numbered columns). This indicates that the in-group effect for minority religions is smaller in those states where the minority religion population is larger. For example, the effect of cultural proximity for minority religions on total credit is 29 percentage points smaller in states with above median minority religion populations. To explore the group heterogeneity dimension holding the group constant, we focus on the subsample of Hindu borrowers and construct a branch-group-quarter panel group- ing Hindu borrowers by their government sanctioned caste classification. Analysis in this panel also allows us to focus on cultural proximity along the caste dimension, holding religion constant. We define the indicator variable SameCaste as equal to one for bor- rowers of caste g in branch b in quarter q when the head officer in the branch belongs

31 to the same caste g. The estimated caste-based in-group effects are presented in Table 10, panel 2. We show the estimated coefficients for SameCaste interacted with indicators for caste equal to SC/ST and General/OBC. This partition of the groups is necessary to explore the effects of group prevalence below, since the 2001 Census does not distinguish OBC from General caste individuals (the interactions for each caste separately are shown in Appendix Table A.3). The point estimates of the in-group effect is significant in al- most every credit outcome for both SC/CT and General/OBC castes. The total credit specification estimates imply that having an in-group officer increases credit by 25% for SC/ST borrowers and by 4% for General/OBC borrowers. As before, when we augment the specification using an interaction with an indicator for whether the state’s fraction of SC/ST population is above the median (0.227), in most specifications the interaction term has the opposite sign of the main effect. Three conclusions arise from this analysis. First, the effect of cultural proximity be- tween members of a group exhibits substantial heterogeneity across group definitions. Cultural proximity has a very large effect on transactions when occurring between mem- bers of a minority religion, or between members of the same backward caste (SC/ST). In contrast, when Hindus are considered as a single group, the effect of cultural proximity on borrowing is weak and does not affect transaction outcomes. Second, much of this heterogeneity is driven by the coarseness of the grouping. Esti- mations based on broad group classifications that collect individuals with distinct cultural backgrounds can fail to capture the existence of in-group preferential treatment even when it is present in smaller partitions of the group. There are strong in-group effects once we look at more homogeneous groups within the Hindu religion, based on government sanc- tioned caste classifications. We show in the Appendix tables A.4 and A.5 that the same occurs for General Class borrowers. The in-group effect on total credit for General Class borrowers is 4% when considered as a single group, but it is much higher once General Class borrowers are partitioned into more homogeneous groups based on Varna member- ship (e.g. 17.4% for Brahmins).

32 Finally, there is a negative correlation between the magnitude of the in-group effect and group size in the population, holding group definition constant. Although there are many potential interpretations for this equilibrium relationship, there are explanations consistent with the existence of information advantages for in-group loans. One potential explanation is that the screening advantage of an in-group officer relative to an out-group one diminishes with the officers’ exposure to the specific cultural traits of other groups. For example, Hindu officers are more likely to be exposed to the cultural traits of Muslims in West Bengal (over 25% Muslim) than in Punjab (1.5% Muslim). As a result, it is likely that the relative advantage of Muslim officers in screening Muslim borrowers is larger in the latter case, consistent with our findings. Explanations related to enforcement and search costs are also consistent with these patterns. For example, Muslim officers may have an advantage in tracking down and censuring Muslim borrowers that default (relative to Hindu officers), but this advantage may diminish if relative the size of the Muslim population is large. (Although this enforcement-based explanation is possible in theory, it requires that in-group officers have an advantage in direct enforcement, a mechanism that is at odds with some of our findings.)

7 Conclusion

In this paper, we have measured the extent of differential treatment in the loan market of those with a shared cultural background. Our empirical context provides a near- ideal setting for assessing differential in-group treatment: since we have data on both lender and borrower group affiliations, we may distinguish between own-group preferences versus differential treatment of minorities. Further, exogenous officer rotation allows us to identify in-group preferences from changes in officer branch assignments. Finally, since we focus on credit markets we may distinguish between explanations based on information, enforcement, and collusion by analyzing loan outcomes. Overall, our findings indicate that better screening and enforcement explain in-group preferential treatment.

33 Our study has a number of implications for theories of discrimination as well as eco- nomic policy. First, we note that the preferential treatment we uncover can itself per- petuate income inequality among minorities. In our context, 74.4% of the officers belong to the General Class category. This implies that the probability that a backward caste borrower (SC, ST, or OBC) will face unfavorable loan conditions is nearly 75%, purely for reasons of cultural affiliation. Further, our findings suggest one possible mechanism through which statistical dis- crimination against minorities can arise. Minorities will be infrequently “matched” with a loan officer of their own group and hence have inferior loan outcomes on average. As a result lenders may form what are ultimately self-confirmatory beliefs about the creditwor- thiness of minorities if they rely on past average group performance to generate lending rules (Kim and Loury, 2009). Finally, our findings have several policy implications. In the Indian context, targeted reservation policies that impose a larger proportion of backward caste officers in regions with a high concentration of backward caste borrowers may improve efficiency and reduce inequality of loan allocation. The reason, however, is different than the preference-based rationales for political reservations (see, for example, Chattopadhyay and Duflo, 2004). Our analysis suggests that reservations may improve contracting outcomes because they reduce information asymmetries between bureaucrats and their communities. Further, policies aimed at reducing cultural differences across groups - for example, by teaching a common language - may lead to improvements in cross-group contracting. However, further research is required to identify which dimensions of cultural heterogeneity have a first-order effect on reducing the ability to exchange information across group boundaries.

References

[1] Agarwal, Sumit and Robert Hauswald, 2010, “Distance and Private Information in Lending,” Review of Financial Studies, 23(7), pages 2757-2788.

34 [2] Aleem, Irfan, 1990, “Imperfect Information, Screening, and the Costs of Informal Lending: A Study of a Rural Credit Market in Pakistan,” The World Bank Economic Review, 4(3), pages 329-349.

[3] Ayres, Ian and Peter Siegelman, 1995, “Race and Gender Discrimination in Bargain- ing for a New Car,” The American Economic Review, 85(3), pages 304-21.

[4] Banerjee, Abhijit, Marianne Bertrand, Saugato Datta, and Sendhil Mullainathan, 2009, “Labor market discrimination in Delhi: Evidence from a field experiment,” Journal of Comparative Economics, 37(1), pages 14-27.

[5] Banerjee, Abhijit, and Esther Duflo, 2010, “Giving Credit Where It Is Due,” Journal of Economic Perspectives, 24(3), pages 61-80.

[6] Banerjee, Abhijit, Shawn Cole and Esther Duflo, 2004, “Banking Reform in India,” India Policy Forum, vol. 1, pages 277-302.

[7] Banerjee, Abhijit, and Kaivan Munshi, 2004, “How Efficiently is Capital Allocated? Evidence from the Knitted Garment Industry in Tiripur,” Review of Economic Stud- ies, 71, pages 19-42.

[8] Becker, Gary, [1957] 1971, “The Economics of Discrimination,” 2nd ed. Chicago: Univ. Chicago Press.

[9] Becker, Gary (1996), “Preferences and Values,” in Becker Gary (ed.), Accounting for Taste, Harvard University Press: Cambridge.

[10] Bandiera, Oriana, Iwan Barankay and Imran Rasul, 2009, “Social Connections and Incentives in the Workplace: Evidence from Personnel Data,” Econometrica, 77(4), pages 1047-94.

[11] Bandiera, Oriana, Iwan Barankay and Imran Rasul, 2005, “Social Preferences and the Response to Incentives: Evidence from Personnel Data,” Quarterly Journal of Economics, 120(3), pages 917-62.

[12] Berger, Allen, Leora Klapper and Gregory Udell, 2001, “The Ability of Banks to Lend to Informationally Opaque Small Businesses,” Journal of Banking and Finance, 25, pages 2127-2167.

[13] Bertrand, Marianne, and Sendhil Mullainathan. 2004, “Are Emily and Greg More Employable Than Lakisha and Jamal? A Field Experiment on Labor Market Dis- crimination,” American Economic Review, 94(4), pages 991-1013.

[14] Beteille, Andre, 1981, The backward classes and the new social order, Oxford Uni- versity Press, Delhi.

35 [15] Beteille, Andre, 1992, The backward classes in contemporary India Delhi, Oxford University Press.

[16] Beteille, Andre, 1996, Caste, class and power. Changing patterns of stratification in a Tanjore village, Oxford University Press, (2nd edn).

[17] Bharat Matrimony. India Matrimonials. http://www.bharatmatrimony.com

[18] Bindu. Kapu surname and gotralu. Kapu sangam forum. Last modified April 29, 2009. Accessed June 20, 2010. http://www.kapusangam.com/forum/index.php?topic=25.0.

[19] Blanchflower, David G., Levine, Phillip B. and David J. Zimmerman, 2003, “Dis- crimination in the Small-Business Credit Market,” The Review of Economics and Statistics, 85(4), pages 930-943.

[20] Buhler, G., 1886, “Sacred Books of the East: The Laws of Manus,” (Vol. XXV).

[21] Burchardi, Konrad B. and Tarek A. Hassan, 2010, “The Economic Impact of Social Ties: Evidence from German Reunification,” Chicago Booth Working Paper.

[22] Chandigarhiya, posting to Khatri Surnames List, UNP, Feb 10, 2010. http://www.unp.co.in/f16/khatri-surnames-68440/

[23] Chattopadhyay, Raghabendra and Esther Duflo, 2004, “Women as Policy Makers: Evidence from a Randomized Policy Experiment in India,” Econometrica, 72(5), pages 1409-1443.

[24] Cohen, Lauren, Andrea Frazzini, and Christopher Malloy, 2008, “The Small World of Investing: Board Connections and Mutual Fund Returns,” Journal of Political Economy, 116(5), pages 951-979.

[25] Cohen, Lauren, Andrea Frazzini, and Christopher Malloy, 2010, “Sell Side School Ties,” forthcoming, Journal of Finance.

[26] Cornell, Bradford, and Ivo Welch, 1996, “Culture, Information, and Screening Dis- crimination,” Journal of Political Economy, 104(3), pages 542-571.

[27] Dahiya, B.S. Jats, the ancient rulers, a clan study. New Delhi: Sterling, 1980.

[28] Dudhane, V.P., 1996, “Maratha Clans.” Last modified Oct 19, 2007. Accessed June 20, 2010. http://bluwiki.com/go/96 Maratha Clans.

[29] Engelberg, Joey, Pengjie Gao and Chris Parsons, 2011, “Friends with Money,” forth- coming, Journal of Financial Economics.

[30] Fafchamps, Marcel, 2000, “Ethnicity and credit in African manufacturing,” Journal of Development Economics, 61(1), pages 205-235.

36 [31] Field, Erica, Matthew Levinson, Rohini Pande, Sujata Visaria, 2008, “Segregation, Rent Control, and Riots: The Economics of Religious Conflict in an Indian City,” American Economic Review P&P, 98(2), pages 505-510.

[32] Fisman, Raymond J., 2003, “Ethnic Ties and the Provision of Credit: Relationship- Level Evidence from African Firms,” The B.E. Journal of Economic Analysis and Policy, Berkeley Electronic Press, 1(1).

[33] Fryer, Roland and Loury, Glenn, 2005, “Affirmative action in winner-take-all mar- kets,” Journal of Economic Inequality, 3(3), pages 263-280.

[34] Fuller, C.J., 1992, The camphor flame. Popular Hinduism and society in India, Princeton, Princeton University Press.

[35] Goldin, Claudia, Rouse, Cecilia, 2000. “Orchestrating impartiality: The impact of blind auditions on female musicians,” American Economic Review, 90(4), pages 715- 741.

[36] Guiso, Luigi, Paola Sapienza and Luigi Zingales, 2009, “Cultural Biases in Economic Exchange?” Quarterly Journal of Economics, 124(3), pages 1095-1131.

[37] Hwang, Byoung-Hyoun and Seoyoung Kim, 2009, “It Pays to Have Friends,” Journal of Financial Economics, 93(1), pages 138-158.

[38] Jackson, Kirabo and Henry Schneider, 2011, “Do Social Connections Reduce Moral Hazard? Evidence from the New York City Taxi Industry,” American Economic Journal: Applied Economics, 3(July), pages 244-267.

[39] Jeevan Sathi. Matrimonials. http://www.jeevansathi.com. Accessed June 24, 2010.

[40] Kamat, V. A List of Brahmin Communities. The Brahmins. Last modified Jan 4 2011. http://www.kamat.com/kalranga/people/brahmins/list.htm.

[41] Kaul, A.. Kashmiri Surnames Database. Kashmir Pundit Surnames. Accessed June 23, 2010. http://kauls.net/surnames/

[42] Khare, R.S 1984, The untouchable as himself. Ideology, identity and pragmatism among the Lucknow Chamars, Cambridge,Cambridge University Press.

[43] Kofi Charles, Kerwin and Guryan, Jonathan, 2008, “Prejudice and Wages: An Em- pirical Assessment of Becker’s The Economics of Discrimination,” Journal of Political Economy, 116(5), pages 773-809.

[44] Kolenda, Pauline 1983, Caste, cult and hierarchy. Essays on the culture of India, New Delhi, Manohar.

37 [45] Kolenda, Pauline, 1986, Caste in India since Independence in D. K. Basu and R. Sissons (eds.), Social and economic development in India Delhi: Sage Publications, pages 106-28.

[46] Ladd, Helen F, 1998, “Evidence on Discrimination in Mortgage Lending,” Journal of Economic Perspectives, 12(2), pages 41-62.

[47] Lerner, Josh and Ulrike Malmedier, 2011, “With a Little Help from My (Random) Friends: Success and Failure in Post-Business School Entrepreneurship,” Manuscript, Harvard University and U.C. Berkeley.

[48] List, John A., 2004, “The Nature and Extent of Discrimination in the Marketplace: Evidence From the Field,” Quarterly Journal of Economics, 119(1), pages 49-89.

[49] Maganti, V., “Brahmins”. Accessed June 25, 2010. http://www.maganti.org/PDFdocs/brahmins.pdf

[50] Maheshwarisamaj. Maheshwari Samaj- Khap. Accessed June 23, 2010. http://www.maheshwarisamaj.com/origin/khap.aspx

[51] Marathas. Kshatriyas and Jainism. Accessed June 25, 2010. http://marathas.tripod.com/kshatrijainism.html

[52] Mian, Atif, 2006, “Distance Constraints: The Limits of Foreign Lending in Poor Economies,” Journal of Finance, 61(3), pages 1465-1505.

[53] Mohanty, P.K., 2004,Encyclopaedia of primitive tribes in India, Volume 2, Kalpaz.

[54] NCBC, Central list of OBCs. Accessed June 20, 2010, http://ncbc.nic.in/backward- classes/index.html.

[55] Norman, Peter, 2003, “Statistical Discrimination and Efficiency,” Review of Eco- nomic Studies, 70(3), pages 615-627.

[56] Parry, Jonathan, 1980, “Ghosts, greed and sin: The occupational identity of the Benares funeral priests,” Man, 15(1), pages 88-111.

[57] Parsons, Christopher, Johan Sulaeman, Michael Yates and Daniel Hamermesh, 2011, “Strike Three: Discrimination, Incentives, and Evaluation,” American Economic Re- view, 101(4), pages 1410-1435.

[58] Petersen, Mitchell and Raghuram Rajan, 1994, “The Benefits of Lending Relation- ships: Evidence from Small Business Data,” Journal of Finance, 49(1), pages 3-37.

[59] Phelps, E., 1972, “The statistical theory of racism and sexism,” American Economic Review, No. 62, pages 659-661.

38 [60] Pope, Devin G., and Justin R. Sydnor, 2010, “What’s in a Picture? Evidence of Discrimination from Prosper.com,” Journal of Human Resources, forthcoming.

[61] Schoar, Antoinette, Raj Kamal Iyer and Sandya Kumar, 2008, “Importance of Ethnic Networks for Business Transactions of the Small Enterprises,” Institute for Financial Management and Research Small Enterprise Finance Centre Working Paper.

[62] Shue, Kelly, 2011, “Executive Networks and Firm Policies: Evidence from the Ran- dom Assignment of MBA Peers,” Chicago Booth Working Paper.

[63] Singh, K.S., Lavanta, B.K., Samanta, D.K., Mandal, S.K., and Vyas, N.N., 1998, People of India-Rajasthan, Popular Prakashan, Part I Volume XXXVIII .

[64] Singh, K.S and Bhanu, B.V., 2004, People of India-Maharashtra, Popular Prakashan, Part II Volume XXX.

[65] Singh, K.S. and Lal, R.B., 2003, People of India-Gujarat, Popular Prakashan, Part I Volume XXII.

[66] Stein, Jeremy C., 2002, “Information Production and Capital Allocation: Decentral- ized versus Hierarchical Firms,” Journal of Finance, 57(5), pages 1891-1921.

[67] Surfindia. Agarwal Caste Agarwal Matrimonials. Accessed June 20, 2010. http://www.surfindia.com/matrimonials/agarwal.html.

[68] Tajfel, H. (1982). Social identity and intergroup relations. Cambridge, England: Cam- bridge University Press.

39 Figure 1: Geographical Distribution of Branches, Weighted by Total Lending

The centers of the circles indicate the location of the branches. The area represents the total amount of lending in the branch in 2002.

40 Figure 2: Average (Relative) Debt by Time since Officer Change

(a) Borrowers that Transition from Out-Group to In-Group Officer, Relative to Borrowers that Transition from Out- Group to Out-Group Officer 2,500

es) 2,000 pe u u R 0 1,500 00 1 (x 1,000 ce n ere f

if 500 D bt e 0 D 0 1 2 3 4 5 6 -6 -5 -4 -3 -2 -1 ge a -500 ver A -1,000 Quarters since Officer Change

(b) Borrowers that Transition from In-Group to Out-Group Officer, Relative to Borrowers that Transition from In-Group to In-Group Officer 40004,000 s) e

pe 2,000 u R 0 0 00

1 0 (x 6 5 4 3 2 1 0 1 2 3 4 5 6 ------e

nc -20002,000 ere f f if D D

bt -4,000 e D e ag r r -60006,000 ve A -8,000 Quarters since Officer Change

The horizontal axis measures time, in quarters, since the officer transition (0 represent the first quarter of the new officer). The vertical axis measures the average debt difference calculated based on a classification of borrowers and officers into five minority religions and four government sanctioned castes (conditional on Hindu). The dashed lines indicate the 95% confidence interval of the mean differences by quarter.

41 Figure 3: Intensive and Extensive Margin Effect Dynamics

(a) Effect of Cultural Proximity on Total Credit

0.30

0.20

ln(Total Debt) Debt) ln(Total 0.10 Estimated in-groupEstimated on effect

0.00 0 1 2 3 4 5 6 7 8 9 10 11 12 Quarters since Officer Arrived in Branch

(b) Effect of Cultural Proximity on Number of Borrowers

0.15

0.10

0.05 ln(Total Debt) Debt) ln(Total

0.00 Estimated in-groupEstimated on effect 0 1 2 3 4 5 6 7 8 9 10 11 12

-0.05 Quarters since Officer Arrived in Branch

The horizontal axis measures time, in quarters, since the officer arrived in the branch (0 represent the first quarter of the new officer). The vertical axis plots the point estimates and 95% confidence interval of the estimated in-group effect by tenure of the officer in the branch (using specification (1) augmented with interactions between SameGroup and a set of indicator variables for the time of the officer in the branch).

42 Table 1: Borrower and Head Officer Composition, by Religion and Caste

Group refers to the religion and caste conditional on Hindu religion the borrower belongs to. There are nine groups: five minority religions and, conditional on Hindu religion, four government sanctioned castes.

Borrowers (%) Head Officers (%)

Panel 1: by Religion Hindu 89.36 93.79 Muslim 6.33 1.84 Christian 1.81 2.06 Sikh 1.95 1.76 Parsi 0.13 0.05 Buddhist 0.19 0.25 Other 0.23 0.25

Panel 2: by Official Caste General 66.66 74.31 SC 10.67 15.68 ST 6.02 5.12 OBC 16.64 4.89

Panel 3: by Varna Brahmin 18.28 23.01 Kshatriya 60.52 43.43 Vaishya 6.59 11.67 Kshatriya/Brahmin 1.72 10.77 Kshatriya/Brahmin/Vaishya 6.76 3.48 Kshatriya/Vaishya 0.41 1.29 Other 5.72 6.35

43 Table 2: Summary Statistics

Panel 1 presents statistics of the panel aggregated at the branch-quarter level. Panel 2 presents statistics of the panel aggregated at the branch-quarter-group level, where group refers to the religion and caste conditional on Hindu religion the borrower belongs to (there are nine groups: five minority religions and, conditional on Hindu religion, four government sanctioned castes). We report, the mean, standard deviation, 1st percentile, median and 99th percentile for all variables.

Mean Std. Dev. p1 p50 p99

Panel 1. Branch-Quarter Statistics, N = 46,753 Total Credit (millions of rupees) 19.25 20.14 1.14 13.52 101.55 # of Borrowers 416.9 362.0 22.0 334.0 1,777.0 # of Different Borrower Religions 3.47 0.72 1.00 4.00 4.00 # of Different Borrower Castes 3.25 1.18 1.00 3.00 6.00 # of Different Borrower Groups (5 minority religions, 4 castes) 5.67 1.49 2.00 6.00 9.00 # of Loan Officers (Including Head Officer) 3.53 4.20 0.00 2.00 16.00 # of Clerks 6.41 7.12 0.00 4.00 31.00

Panel 2. Group-Branch-Quarter Statistics, N = 339,366 Sum Debt (1,000s of rupees) 2006.0 7068.0 0.0 42.0 31307.0 Std. Dev. Debt (1,000s of rupees) 47.0 121.0 0.0 10.0 395.0 IQR Debt (1,000s of rupees) 31.0 131.0 0.0 0.0 380.0 Dummy = 1 if Debt > 0 0.578 0 0 1 1 Sum Collateral (1,000s of rupees) 12,267 792,107 0 100 113,062 Std. Dev. Collateral (1,000s of rupees) 745 70,603 0 18 1,163 IQR Collateral (1,000s of rupees) 82.00 1100.00 0.00 0.00 865.00 # of Borrowers 43.40 127.00 0.00 2.00 544.00 Fraction of Borrowers with Over 60 Days Late 0.081 0.172 0.000 0.000 1.000 Fraction of Debt with Over 60 Days Late 0.020 0.116 0.000 0.000 1.000 SameGroup 0.11 0.31 0.00 0.00 1.00

44 Table 3: Effect of Cultural Proximity on Credit: LATE and ATE Bounds

In this table we report the estimated effect of cultural proximity on the log debt amount (column 1), log number of borrowers (column 4) using specification (1). The unit of analysis is a branch-group-quarter, where group is defined by combining religion and caste based measures of cultural proximity (five minority religions and four government designated castes conditional on Hindu religion). The variable SameGroup is an indicator denoting that borrowers and the branch manager are of the same group. Standard errors are clustered at the branch level. Column 7 shows the a linear probability model estimate of the effect of proximity on receiving any credit. Columns 2, 3, 5, and 6, presents the upper and lower bounds of the local average treatment effect, estimated as in Lee (2009): specification (1) is estimated after trimming the upper and lower 1.5 percentiles (estimated treatment effect on the linear probability model) of the observations in the treatment group (SameGroup = 1) by the dependent variable. * significant at 10%; ** significant at 5%; *** significant at 1%

Dependent Variable ln(Total Debt) ln(# of Borrowers) Dummy = 1 if Total Debt > 0 Trim Percentile None 1.5 98.5 None 1.5 98.5 None (1) (2) (3) (4) (5) (6) (7)

SameGroup 0.1856*** 0.1826*** 0.2293*** 0.0619*** 0.0596*** 0.0785*** 0.0155*** (0.020) (0.020) (0.019) (0.013) (0.013) (0.013) (0.004) Branch-Group Fixed Effects Yes Yes Yes Yes Yes Yes Yes 45 Quarter Dummies Yes Yes Yes Yes Yes Yes Yes Observations 225,543 224,900 224,900 234,451 233,806 233,923 385,293 R-squared 0.883 0.881 0.884 0.947 0.946 0.947 0.809 Table 4: Identification Test: Saturated Specifications

In this table we report the estimated effect of cultural proximity on lending outcomes using the specifica- tion (1) augmented with branch-quarter, group-quarter, State-group-quarter and district-group-quarter dummies. The unit of analysis is a branch-group-quarter, where group is defined by combining religion and caste based measures of cultural proximity. The variable SameGroup is an indicator denoting that borrowers and the branch manager are of the same group. Standard errors are clustered at the branch level. * significant at 10%; ** significant at 5%; *** significant at 1%

Dependent Variable ln(Total Debt) ln(# of Borrowers) Dummy = 1 if Total Debt > 0 (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12)

SameGroup 0.1711*** 0.1734*** 0.1809*** 0.1800*** 0.0580*** 0.0603*** 0.0612*** 0.0556*** 0.0148*** 0.0140*** 0.0129*** 0.0116** (0.021) (0.021) (0.019) (0.022) (0.014) (0.014) (0.012) (0.014) (0.004) (0.004) (0.004) (0.005) Demeaned: Branch-Group Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Dummies: Branch-Quarter Yes Yes Yes Yes Yes Yes Group-Quarter Yes Yes Yes State-Group-Quarter Yes Yes Yes District-Group-Quarter Yes Yes Yes Observations 225,543 225,543 220,900 221,187 234,451 234,451 229,584 229,885 385,293 385,293 375,927 376,377

46 R-squared 0.317 0.323 0.115 0.339 0.415 0.432 0.169 0.390 0.143 0.147 0.032 0.200 Table 5: Robustness Test: Alternate Group Definition Based on Surnames (Varnas)

In this table we report the estimated effect of cultural proximity on loan outcomes using specification (1). The unit of analysis is a branch-group-quarter, where group is defined by Varna, the caste system that was prevalent in ancient India. Individuals are assigned to Varnas using a surname-matching algorithm. Individuals classified by the algorithm into more than one Varna are allocated to three mixed groups: Kshatriya-Brahmin, Kshatriya-Brahmin-Vaishya, and Kshatriyas-Vaishyas. Individuals that are not in the General Class category are excluded, because the algorithm cannot correctly identify individuals from the Shudra Varna. The variable SameGroup is an indicator denoting that borrowers and the branch manager are of the same group. Standard errors are clustered at the branch level. * significant at 10%; ** significant at 5%; *** significant at 1%

Dependent Variable ln(Total Debt) ln(# of Borrowers) Dummy = 1 if Total Debt > 0 (1) (2) (3)

SameGroup (Varna) 0.1407*** 0.0231 0.0226** (0.035) (0.015) (0.009) Branch-Group Fixed Effects Yes Yes Yes Quarter Dummies Yes Yes Yes Observations 67,202 70,760 108,947 R-squared 0.863 0.950 0.831

47 Table 6: Intensive and Extensive Margins: Existing and First Time Borrowers

In this table we report the estimated effect of cultural proximity on lending patterns (specification (1)) separately for existing borrowers (Panel 1) and first time borrowers (Panel 2). Existing borrowers are those that obtained credit at any time in our sample prior to the arrival of the current officer in charge of the branch. First time borrowers receive their first credit from the Bank under the current officer. The unit of analysis is a branch-group-quarter, where group is defined by combining religion and caste based measures of cultural proximity. The variable SameGroup is an indicator denoting that borrowers and the branch manager are of the same group. Standard errors are clustered at the branch level. * significant at 10%; ** significant at 5%; *** significant at 1%

Dependent Variable ln(Total Debt) ln(# of Borrowers) Dummy = 1 if Total Debt > 0 (1) (2) (3)

Panel 1. Borrowers that had obtained Credit from Bank prior to Officer’s Arrival SameGroup 0.1158*** 0.0029 0.0162*** (0.026) (0.017) (0.005) Branch-Group Fixed Effects Yes Yes Yes Quarter Dummies Yes Yes Yes Observations 196,928 209,507 397,923 R-squared 0.828 0.905 0.820

Panel 2. Borrowers that Obtain Bank Credit for the First Time with the Current Officer SameGroup 0.1693*** 0.1327*** 0.0107 (0.030) (0.023) (0.007) Branch-Group Fixed Effects Yes Yes Yes Quarter Dummies Yes Yes Yes Observations 184,386 190,128 397,035 R-squared 0.786 0.841 0.732

48 Table 7: Effect of Cultural Proximity on Loan Repayment

In this table we report the estimated effect of cultural proximity on loan repayment using specification (1). The unit of analysis is a branch-group-quarter, where group is defined by combining religion and caste based measures of cultural proximity. Repayment is measured as the fraction of loans (columns 1 through 3) and fraction of debt (columns 4 through 6) that is more than 60 days overdue. The variable SameGroup is an indicator denoting that borrowers and the branch manager are of the same group. Average repayment performance is included as the dependent variable contemporaneously in columns 1 and 4, and with a one-year lead in the rest. Columns 5 and 6 include the one year lead of the SameGroup variable in the regression. Standard errors are clustered at the branch level. * significant at 10%; ** significant at 5%; *** significant at 1%.

Dependent Variable Fraction of Borrowers Late Over 60 Days Fraction of Debt Late Over 60 Days Contemporaneous One Year Forward Contemporaneous One Year Forward (1) (2) (3) (4) (5) (6)

Panel 1. All Borrowers SameGroup -0.0057*** -0.0097*** -0.0080** -0.0029*** -0.0054*** -0.0055*** (0.002) (0.003) (0.003) (0.001) (0.001) (0.002) SameGroup One Year Forward -0.0042 -0.0008 (0.003) (0.001) Branch-Group and Quarter Dummies Yes Yes Yes Yes Yes Yes 49 Observations 225,543 210,906 167,733 225,543 210,906 167,733 R-squared 0.339 0.558 0.586 0.359 0.580 0.610

Panel 2. Subsample of Borrowers that had obtained Credit from Bank prior to Officer’s Arrival SameGroup -0.0041* -0.0066* -0.0046 -0.0026** -0.0047** -0.0051** (0.002) (0.004) (0.004) (0.001) (0.002) (0.002) SameGroup One Year Forward -0.0045 -0.0015 (0.004) (0.002) Branch-Group and Quarter Dummies Yes Yes Yes Yes Yes Yes Observations 209,114 194,577 144,950 196,928 184,581 144,950 R-squared 0.377 0.586 0.622 0.415 0.619 0.646

Panel 3. Subsample of Borrowers that Obtain Credit from the Bank for the First Time with the Current Officer SameGroup -0.0032* -0.0091** -0.0080* -0.0012 -0.0045* -0.0040 (0.002) (0.004) (0.004) (0.001) (0.003) (0.003) SameGroup One Year Forward -0.0033 -0.0025 (0.004) (0.002) Branch-Group and Quarter Dummies Observations 134,227 134,227 132,788 134,227 134,227 132,788 R-squared 0.322 0.537 0.538 0.330 0.545 0.546 Table 8: Effect of Cultural Proximity on the Dispersion of Lending and Collateral

In this table we report the estimated effect of cultural proximity on the log standard deviation and inter- quartile range of debt (columns 1 and 2), and the log total, standard deviation, and inter-quartile range of collateral (columns 3, 4 and 5) using specification (1). The unit of analysis is a branch-group-quarter, where group is defined by combining religion and caste based measures of cultural proximity (five minority religions and four government designated castes conditional on Hindu religion). The variable SameGroup is an indicator denoting that borrowers and the branch manager are of the same group. Standard errors are clustered at the branch level. * significant at 10%; ** significant at 5%; *** significant at 1%

Dependent Variable ln(Total Collateral) ln(Average Collateral/Loan) ln(Std. Dev. Debt) ln(IQR Debt) (1) (2) (3) (4)

Panel 1. All Borrowers SameGroup 0.1364*** -0.0474*** 0.1832*** 0.0879*** (0.019) (0.012) (0.018) (0.016) Branch-Group and Quarter Dummies Yes Yes Yes Yes Observations 232,126 223,621 197,094 195,321 R-squared 0.916 0.583 0.712 0.668

Panel 2. Subsample of Borrowers that had obtained Credit from Bank prior to Officer’s Arrival SameGroup 0.0739*** -0.0493*** 0.1597*** 0.0722*** 50 (0.024) (0.015) (0.022) (0.021) Branch-Group and Quarter Dummies Yes Yes Yes Yes Observations 207,412 195,360 170,212 165,297 R-squared 0.889 0.584 0.709 0.642

Panel 3. Subsample of Borrowers that Obtain Credit from the Bank for the First Time with the Current Officer SameGroup 0.1859*** 0.0144 0.1197*** 0.0298 (0.033) (0.014) (0.024) (0.022) Branch-Group and Quarter Dummies Yes Yes Yes Yes Observations 187,396 182,005 155,096 154,258 R-squared 0.808 0.499 0.640 0.586 Table 9: Heterogeneity: Branch Density in District and Branch Size

In this table we report the heterogeneity of the estimated effect of cultural proximity on lending outcomes using specification (1) across districts of different branch density, defined as the number of branches per 1,000 inhabitants (panel 1), and by number of officers in the branch (panel 2). The unit of analysis is a branch-group-quarter, where group is defined by combining religion and caste based measures of cultural proximity. The variable SameGroup is an indicator denoting that borrowers and the branch manager are of the same group. Its interactions with an indicator for whether the branch is located in a district with the second, third or fourth quartiles of branch density (panel 1), and with two dummies for whether the branch has two officers, and more than two officers (panel 2). Standard errors are clustered at the branch level. * significant at 10%; ** significant at 5%; *** significant at 1%

Dependent Variable ln(Total Debt) ln(# of Borrowers) Dummy = 1 if ln(IQR Debt) Fraction of Debt Total Debt > 0 Late Over 60 Days (1) (2) (3) (4) (5)

Panel 1: Effect Heterogeneity by Branch Density in the District SameGroup 0.2239*** 0.0948*** 0.0069 0.1033*** -0.0092*** (0.049) (0.034) (0.009) (0.036) (0.003) SameGroup × -0.0582 -0.0238 0.0187 -0.0542 0.0057 Branches per 1000 inhabitant in District = quartile 2 (0.073) (0.049) (0.015) (0.051) (0.004) SameGroup × -0.0346 -0.0265 0.0050 -0.0363 0.0086*** 51 Branches per 1000 inhabitant in District = quartile 3 (0.060) (0.042) (0.011) (0.047) (0.003) SameGroup × -0.0436 -0.0452 0.0082 -0.0024 0.0078** Branches per 1000 inhabitant in District = quartile 4 (0.059) (0.040) (0.012) (0.046) (0.003) Branch-Group Fixed Effects Yes Yes Yes Yes Yes District Quartile-Quarter Dummies Yes Yes Yes Yes Yes Observations 214,436 222,508 362,736 185,819 214,436 R-squared 0.884 0.947 0.811 0.671 0.360

Panel 2: Effect Heterogeneity by Number of Officers in the Branch SameGroup 0.2300*** 0.0600** 0.0252*** 0.1309*** -0.0042*** (0.041) (0.025) (0.008) (0.031) (0.002) SameGroup × -0.0214 0.0130 -0.0031 -0.0390 0.0021 # of Officers in branch = 2 (0.052) (0.036) (0.012) (0.042) (0.002) SameGroup × -0.0870* -0.0019 -0.0197** -0.0828** 0.0017 # of Officers in branch > 2 (0.050) (0.032) (0.010) (0.039) (0.002) Branch-Group Fixed Effects Yes Yes Yes Yes Yes Branch Size-Quarter Dummies Yes Yes Yes Yes Yes Observations 220,600 229,110 377,248 190,844 220,600 R-squared 0.883 0.946 0.809 0.671 0.354 Table 10: Heterogeneity by Religion, Caste, and Group Size

In this table we report the estimated effect of cultural proximity through religion (panel 1) and caste (panel 2) on lending patterns using specification (1). The unit of analysis is a branch-group-quarter, where group is defined by religion in panel 1 (Hindu, Muslim, Christian, Sikh, Parsi, and Buddhist), and by government designated caste in the subsample of Hindu borrowers in panel 2 (General Class/SC/ST/OBC). The variable SameReligion (SameCaste) is an indicator denoting that borrowers and the branch manager are of the same religion (caste). The variables Hindu and Minority Religion are dummies equal to one if the borrower is of the corresponding religion. The variables SC/ST and OBC/General are dummies equal to one if the borrower belongs to the corresponding caste. OBC and General Class individuals are grouped together because these two groups are not distinguished in the 2001 India Census. Standard errors are clustered at the branch level. * significant at 10%; ** significant at 5%; *** significant at 1%

Dependent Variable ln(Total Debt) ln(# of Borrowers) Dummy = 1 if ln(IQR Debt) Fraction of Debt Total Debt > 0 Late Over 60 Days (1) (2) (3) (4) (5) (6) (7) (8) (9) (10)

Panel 1: Effect Heterogeneity by Religion SameReligion × Hindu 0.0148 0.0153 -0.0183 -0.0292 0.0003 -0.0005 0.0387 0.0266 -0.0000 0.0000 (0.025) (0.032) (0.024) (0.033) (0.002) (0.003) (0.029) (0.036) (0.000) (0.000) SameReligion × Minority Religion 0.4307*** 0.5981*** 0.1497*** 0.1412** 0.0643*** 0.1032*** 0.1927*** 0.3871*** -0.0085** -0.0076 (0.066) (0.112) (0.041) (0.065) (0.015) (0.025) (0.055) (0.108) (0.004) (0.006) 52 SameReligion × Hindu -0.0035 0.0190 0.0015 0.0238 -0.0001 × Large Minority State (0.049) (0.048) (0.005) (0.057) (0.000) SameReligion × Minority Religion -0.2942** 0.0173 -0.0711** -0.3331*** -0.0016 × Large Minority State (0.135) (0.083) (0.030) (0.119) (0.008) Branch-Group Fixed Effects Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Quarter Dummies Yes Yes Yes Yes Yes State-Quarter Dummies Yes Yes Yes Yes Yes Observations 130,849 130,849 137,130 137,130 264,372 264,372 109,507 109,507 130,849 130,849 R-squared 0.924 0.924 0.973 0.973 0.814 0.814 0.687 0.687 0.369 0.369

Panel 2: Effect Heterogeneity by Government Sanctioned Caste (Hindu subsample) SameCaste × SC/ST 0.2514*** 0.2455*** 0.1066*** 0.1617*** 0.0241* 0.0409** 0.1123** 0.0759 -0.0074** -0.0088** (0.060) (0.066) (0.038) (0.047) (0.013) (0.017) (0.056) (0.068) (0.003) (0.004) SameCaste × OBC/General 0.0414** 0.0461* 0.0419** 0.0420** 0.0001 -0.0005 0.0040 0.0074 -0.0003 -0.0001 (0.021) (0.024) (0.017) (0.020) (0.003) (0.004) (0.018) (0.021) (0.001) (0.001) SameCaste × SC / ST 0.0116 -0.1769** -0.0556** 0.1136 0.0042 × Large SC / ST State (0.141) (0.079) (0.023) (0.118) (0.005) SameCaste × OBC/General -0.0164 -0.0066 0.0003 -0.0091 -0.0008 × Large SC / ST State (0.047) (0.037) (0.006) (0.039) (0.001) Branch-Group Fixed Effects Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Quarter Dummies Yes Yes Yes Yes Yes State-Quarter Dummies Yes Yes Yes Yes Yes Observations 150,250 150,250 152,719 152,719 176,248 176,248 140,633 140,633 150,250 150,250 R-squared 0.889 0.890 0.936 0.936 0.734 0.734 0.689 0.689 0.334 0.335 Table A.1: Empirical Officer Group Transition Frequencies

In this table we report the empirical branch officer transition probabilities, by officer religion (panel 1), caste (panel 2), and Varna (panel 3), and the results of the χ2 test of equality between empirical and the theoretical transition probabilities officers are randomly allocated to branches. * significant at 10%; ** significant at 5%; *** significant at 1%.

Panel 1: Empirical Distribution of Branch Transitions, by Religion Fom Religion: Hindu Muslim Christian Sikh Parsi Buddist Others To religion: Hindu 87.173% 1.712% 2.423% 1.454% 0.032% 0.291% 0.129% Muslim 1.842% 0.097% 0.032% 0.032% Christian 2.003% 0.065% 0.452% 0.032% Sikh 1.389% 0.323% 0.032% Parsi 0.032% Buddist 0.194% 0.032% Others 0.162% 0.065%

Panel 2: Empirical Distribution of Branch Transitions, by Caste From Caste: SC ST OBC General To caste SC 3.70% 0.70% 0.73% 9.87% ST 0.63% 0.80% 0.35% 2.48% OBC 0.73% 0.31% 0.70% 3.52% General 8.37%* 2.34% 3.73% 61.03%***

Panel 3: Empirical Distribution of Branch Transitions, by Varna

From Varna: Brahmin Kshatriya Vaishya Multiple Matches Not Matched To Varna: Brahmin 3.01% 5.37% 1.55% 2.03% 5.40% Kshatriya 5.37% 9.57% 2.76% 3.62% 9.62% Vaishya 1.55% 2.76% 0.79% 1.04% 2.77% Multiple Matches 2.03% 3.62% 1.04% 1.37% 3.64% Not Matched 5.40% 9.62% 2.77% 3.64% 9.68%

53 Table A.2: Effect of Cultural Proximity on Loan Repayment, Alternative Default Horizons

In this table we report the estimated effect of cultural proximity on loan repayment using specification (1). The unit of analysis is a branch-group-quarter, where group is defined by combining religion and caste based measures of cultural proximity. Repayment is measured as the fraction of loans (columns 1 through 3) and fraction of debt (columns 4 through 6) that is more than 30 or 90 days overdue (panels 1 and 2 respectively). The variable SameGroup is an indicator denoting that borrowers and the branch manager are of the same group. Average repayment performance is included as the dependent variable contemporaneously in columns 1 and 4, and with a one-year lead in the rest. Columns 5 and 6 include the one year lead of the SameGroup variable in the regression. Standard errors are clustered at the branch level. * significant at 10%; ** significant at 5%; *** significant at 1%.

Dependent Variable Fraction of Borrowers Late Fraction of Debt Late Contemporaneous One Year Forward Contemporaneous One Year Forward (1) (2) (3) (4) (5) (6)

Panel 1. Late Over 30 Days SameGroup -0.0057*** -0.0098*** -0.0080** -0.0029*** -0.0054*** -0.0054*** (0.002) (0.003) (0.003) (0.001) (0.001) (0.002) SameGroup One Year Forward -0.0042 -0.0008 (0.003) (0.001) Branch-Group and Quarter Dummies Yes Yes Yes Yes Yes Yes 54 Observations 225,543 210,906 167,733 225,543 210,906 167,733 R-squared 0.340 0.558 0.587 0.359 0.581 0.611

Panel 2. Late Over 90 Days SameGroup -0.0057*** -0.0097*** -0.0080** -0.0029*** -0.0054*** -0.0055*** (0.002) (0.003) (0.003) (0.001) (0.001) (0.002) SameGroup One Year Forward -0.0041 -0.0008 (0.003) (0.001) Branch-Group and Quarter Dummies Observations 225,543 210,906 167,733 225,543 210,906 167,733 R-squared 0.339 0.558 0.586 0.359 0.580 0.610 Table A.3: Heterogeneity by Religion

In this table we report the estimated effect of cultural proximity through religion on lending patterns using specification (1). The unit of analysis is a branch-group-quarter, where group is defined by religion (Hindu, Muslim, Christian, Sikh, Parsi, and Buddhist). The variable SameReligion is an indicator denoting that borrowers and the branch manager are of the same religion. The variables Hindu, Muslim, Christian, Sikh, Parsi, Buddhist, and Minority Religion are dummies equal to one if the borrower is of the corresponding religion. Standard errors are clustered at the branch level. * significant at 10%; ** significant at 5%; *** significant at 1%

Dependent Variable ln(Total Debt) ln(# of Borrowers) Dummy = 1 if ln(IQR Debt) Fraction of Debt Total Debt > 0 Over 60 Days Late (1) (2) (3) (4) (5)

SameReligion × Hindu 0.0149 -0.0182 0.0003 0.0386 -0.0000 (0.025) (0.024) (0.002) (0.029) (0.000) SameReligion × Muslim 0.6059*** 0.1538** 0.0285** 0.2244** -0.0193** (0.129) (0.078) (0.011) (0.098) (0.009) SameReligion × Christian 0.3008*** 0.1079** 0.0803*** 0.1907** -0.0020 (0.095) (0.054) (0.026) (0.088) (0.004) SameReligion × Sikh 0.3535*** 0.1410* 0.0648 0.1860 -0.0055 (0.127) (0.084) (0.040) (0.116) (0.005) 55 SameReligion × Parsi 0.8076*** 0.0286 0.1340 1.1044*** -0.0005 (0.185) (0.145) (0.130) (0.011) (0.003) SameReligion × Buddhist 0.5573* 0.5312* 0.1427** -0.0159 -0.0015 (0.295) (0.281) (0.061) (0.195) (0.002) Branch-Religion Fixed Effects Yes Yes Yes Yes Yes Quarter Dummies Yes Yes Yes Yes Yes Observations 138,619 145,350 280,500 115,983 138,619 R-squared 0.924 0.974 0.815 0.686 0.370 Table A.4: Heterogeneity by Caste, Hindu Borrower Subsample

In this table we report the estimated effect of cultural proximity through caste on lending patterns using specification (1). The unit of analysis is a branch-group-quarter, where group is defined by govern- ment designated caste, and conditioning on Hindu borrowers (General Class/SC/ST/OBC). The variable SameCaste is an indicator denoting that borrowers and the branch manager are of the same caste. The variables General Class, SC, ST, OBC, and SC/ST/OBC are dummies equal to one if the borrower be- longs to the corresponding caste. Standard errors are clustered at the branch level. * significant at 10%; ** significant at 5%; *** significant at 1%

Dependent Variable ln(Total Debt) ln(# of Borrowers) Dummy = 1 if ln(IQR Debt) Fraction of Debt Total Debt > 0 Over 60 Days Late (1) (2) (3) (4) (5)

SameCaste × General 0.0404** 0.0252 0.0036** 0.0226 -0.0003 (0.020) (0.017) (0.002) (0.017) (0.000) SameCaste × SC 0.1883*** 0.0258 0.0049 0.1009*** -0.0007 (0.039) (0.027) (0.006) (0.034) (0.001) SameCaste × ST 0.2500** 0.0129 0.0411* 0.2337** -0.0075* (0.099) (0.044) (0.023) (0.093) (0.004) SameCaste × OBC 0.2941*** 0.2130*** 0.0119 0.0371 -0.0076** (0.067) (0.054) (0.013) (0.063) (0.004) 56 Branch-Caste Fixed Effects Yes Yes Yes Yes Yes Quarter Dummies Yes Yes Yes Yes Yes Observations 159,761 162,350 187,000 149,678 159,761 R-squared 0.890 0.935 0.736 0.689 0.337 Table A.5: Heterogeneity by Varna

In this table we report the estimated effect of cultural proximity through Varna on lending patterns using specification (1). The unit of analysis is a branch-group-quarter, where group is defined by Varna, the caste system that was prevalent in ancient India. Individuals are assigned to Varnas us- ing a surname-matching algorithm. Individuals classified by the algorithm into more than one Varna are allocated to three mixed groups: Kshatriya-Brahmin, Kshatriya-Brahmin-Vaishya, and Kshatriyas- Vaishyas. The variable SameVarna is an indicator denoting that borrowers and the branch manager are of the same Varna. The variables Brahmin, Kshatriya, Vaishya, B-K (Brahmin-Kshatriya), B-K-V (Kshatriya-Brahmin-Vaishya), and K-V (Kshatriyas-Vaishyas) are dummies equal to one if the borrower belongs to the corresponding Varna. Standard errors are clustered at the branch level. * significant at 10%; ** significant at 5%; *** significant at 1%

Dependent Variable ln(Total Debt) ln(# of Borrowers) Dummy = 1 if ln(IQR Debt) Fraction of Debt Total Debt > 0 Over 60 Days Late (1) (2) (3) (4) (5)

SameVarna × Brahmin 0.1744** 0.0231 0.0256* 0.0814 -0.0091* (0.081) (0.039) (0.015) (0.068) (0.005) SameVarna × Kshatriya 0.0213 0.0310 -0.0028 -0.0042 0.0015 (0.043) (0.025) (0.002) (0.040) (0.001) SameVarna × Vaishya 0.1080 -0.0053 0.0478* 0.0702 0.0015 57 (0.136) (0.054) (0.025) (0.131) (0.009) SameVarna × B-K 0.5148*** 0.0064 0.0201 0.6145*** 0.0073 (0.152) (0.073) (0.043) (0.236) (0.024) SameVarna × B-K-V 0.4229*** 0.0159 0.0152 0.4887** -0.0120 (0.157) (0.068) (0.041) (0.208) (0.016) SameVarna × K-V 1.3311** 0.1428 0.1863* 1.6186*** -0.0126 (0.523) (0.136) (0.098) (0.558) (0.008) Branch-Varna Fixed Effects Yes Yes Yes Yes Yes Quarter Dummies Yes Yes Yes Yes Yes Observations 67,202 70,661 108,947 57,458 67,202 R-squared 0.864 0.950 0.831 0.668 0.404